Review

The Evolution of Artificial Intelligence in Medical Imaging:

From Computer Science to Machine and Deep Learning
Michele Avanzo 1, * , Joseph Stancanello 2 , Giovanni Pirrone 1 , Annalisa Drigo 1 and Alessandra Retico 3

1 Medical Physics Department, Centro di Riferimento Oncologico di Aviano (CRO) IRCCS, 33081 Aviano, Italy;
giovanni.pirrone@cro.it (G.P.); adrigo@cro.it (A.D.)
2 Elekta SA, 92100 Boulogne-Billancourt, France; joseph.stancanello@elekta.com
3 National Institute for Nuclear Physics (INFN), Pisa Division, 56127 Pisa, Italy; alessandra.retico@pi.infn.it
* Correspondence: mavanzo@cro.it

Simple Summary: Artificial intelligence, now one of the most promising frontiers of medicine, has
a long and tumultuous history punctuated by successes and failures. One of its successes was its
application to medical images. We reconstruct the timeline of the advancements in this field, from
its origins in the 1940s before crossing medical images to early applications of machine learning to
radiology, to the present era where artificial intelligence is revolutionizing radiology.

Abstract: Artificial intelligence (AI), the wide spectrum of technologies aiming to give machines
or computers the ability to perform human-like cognitive functions, began in the 1940s with the
first abstract models of intelligent machines. Soon after, in the 1950s and 1960s, machine learning
algorithms such as neural networks and decision trees ignited significant enthusiasm. More recent
advancements include the refinement of learning algorithms, the development of convolutional
neural networks to efficiently analyze images, and methods to synthesize new images. This renewed
enthusiasm was also due to the increase in computational power with graphical processing units
and the availability of large digital databases to be mined by neural networks. AI soon began to
be applied in medicine, first through expert systems designed to support the clinician’s decision
Citation: Avanzo, M.; Stancanello, J.; and later with neural networks for the detection, classification, or segmentation of malignant lesions
Pirrone, G.; Drigo, A.; Retico, A. The in medical images. A recent prospective clinical trial demonstrated the non-inferiority of AI alone
Evolution of Artificial Intelligence in compared with a double reading by two radiologists on screening mammography. Natural language
Medical Imaging: From Computer processing, recurrent neural networks, transformers, and generative models have both improved
Science to Machine and Deep the capabilities of making an automated reading of medical images and moved AI to new domains,
Learning. Cancers 2024, 16, 3702. including the text analysis of electronic health records, image self-labeling, and self-reporting. The
https://doi.org/10.3390/ availability of open-source and free libraries, as well as powerful computing resources, has greatly
cancers16213702
facilitated the adoption of deep learning by researchers and clinicians. Key concerns surrounding AI
Academic Editors: David Wong and in healthcare include the need for clinical trials to demonstrate efficacy, the perception of AI tools
Jason Li as ‘black boxes’ that require greater interpretability and explainability, and ethical issues related
to ensuring fairness and trustworthiness in AI systems. Thanks to its versatility and impressive
Received: 27 September 2024
results, AI is one of the most promising resources for frontier research and applications in medicine,
Revised: 26 October 2024
Accepted: 29 October 2024
in particular for oncological applications.
Published: 1 November 2024
Keywords: artificial intelligence; medical imaging; neural networks; machine learning; deep learning

Copyright: © 2024 by the authors.

Licensee MDPI, Basel, Switzerland. 1. Introduction
This article is an open access article
Artificial intelligence (AI) permeated medicine slowly but steadily, at first through
distributed under the terms and
seminal works and then with the first commercial systems, until the present day, when
conditions of the Creative Commons
Attribution (CC BY) license (https://
AI now represents one of the most promising frontiers of medicine. Researchers have
creativecommons.org/licenses/by/
shown that AI can perform a wide range of tasks in medical imaging [1], with recent
4.0/). prospective clinical trials indicating that AI achieves performance levels comparable to

Cancers 2024, 16, 3702. https://doi.org/10.3390/cancers16213702 https://www.mdpi.com/journal/cancers

Cancers 2024, 16, 3702 2 of 23

humans in diagnostic tasks [2]. However, the widespread adoption of AI in medicine

remains challenging, as it requires ensuring its safe and ethical use [3,4].
In this narrative review, we will describe the history of the development of AI, from
the first conceptualizations of learning machines in the 1940s to the modern refinements
regarding neural networks, which allow for the successful usage of AI in most, if not all,
human disciplines. We will also describe the advancements in the use of AI in medicine, from
the first expert systems to modern applications of neural networks in imaging, in particular for
oncological applications. The review consists of three main sections: in the first, we describe
the works of the pioneers of AI in the 20th century, a time when AI was not interested in
nor capable of analyzing medical images. The second section describes an era when the
first AI-based classifications of imaging findings were attempted, with the first widely used
machine learning (ML) algorithms. In the present era, medical imaging is being transformed
by the endless possibilities offered by a large spectrum of neural network architectures.

2. AI Before Meeting Medical Imaging: From the Origins to Expert Systems

AI is an umbrella term covering a wide spectrum of technologies aiming to give machines
or computers the ability to perform human-like cognitive functions such as learning, problem-
solving, and decision-making. At its beginning, the goal of AI was to imitate the human mind or,
in the words of Frank Rosemblatt, to make a computer “able to walk, talk, see, write, reproduce
itself and be conscious of its existence” [5]. However, it was soon understood that AI could
achieve better results at well-defined, specific tasks such as playing checkers [6], to the level of
surpassing the performance of humans, e.g., the computer Deep Blue defeating former world
chess champion Garry Kasparov in 1997 [7]. In this section, we describe this early phase.

2.1. Prehistory of AI
The idea of inanimate objects being able to complete tasks that are usually performed
by humans and require “intelligence” dates back to ancient times [8]. The history of AI
started with a group of great visionaries and scientists in the 1900s, including Alan Turing
(London, 1912–Manchester, 1954, Figure 1a), one of the fathers of modern computers. He
devised an abstract computer called the Turing machine, a concept of paramount importance
in modern informatics, as any modern computer device is thought to be a special case of
a Turing machine [9]. He also worked on a device, the Bombe, at Bletchley Park (75 km
northwest of London), which involved iteratively reducing the space of solutions from a set of
message transcriptions to discover the decryption key of enemy messages during World War
II [10]. This process had some resemblance to ML, which hypotheses a model from a set of
observations [11]. A timeline of the origin and development of AI, starting from the Turing
machine to the triumph of artificial neural networks (ANNs), is provided in Figure 2.
In a public lecture held in 1947, Turing first mentioned the concept of a “machine
that can learn from experience” [12] and posed the question “can machine think?” in his
seminal paper entitled “Computing machinery and intelligence” [13]. In this paper, the
“imitation game”, also referred to as the Turing test, was proposed to define if a machine
can think. In this test, a human interrogates another human and the machine alternatively.
If it is not possible to distinguish the machine from the human based on the answers then
the machine passes the test and is considered able to think.
Turing discussed strategies for achieving a thinking machine by programming and
learning. He likened the learning process to that of a child being educated by an adult
who provides positive and negative examples [11,14]. From its very beginning, different
branches of AI emerged. Symbolic AI searches for the proper rule (e.g., IF-THEN) to apply
to the problem at hand by testing/simulating all the possible rules, like in a chess game,
without training [15,16]. On the other hand, ML is characterized by a training phase, where
the machine analyses a set of data, builds a model, and measures model performance
through a function called goal or cost function [17]. The term ML was introduced by Arthur
L. Samuel (Emporia, USA, 1901–Stanford, USA, 1990), who developed the first machine
able to learn to play checkers [6]. The dawn of AI is considered the summer conference at
Cancers 2024, 16, 3702 3 of 23

Dartmouth College (Hanover, NH, USA) in 1956 [18]. At the meeting, “artificial intelligence”
was defined by John McCarthy (Boston, USA, 1927–Stanford, USA, 2011) as “the science
and engineering of making intelligent machines”. This definition, as well as the implicit
definition of AI in the imitation game, escapes the cumbersome issue of defining what
intelligence is [19], making the goals and boundaries of the science of AI blurry. For instance,
in the early years of AI, research clearly targeted computers that could have performance
Cancers 2024, 16, x FOR PEER REVIEW
comparable with those of the human mind (”strong AI”). In later years, the AI community 3 of 24
shifted its aim to more limited realistic tasks, like solving practical problems and 3carrying
Cancers 2024, 16, x FOR PEER REVIEW of 24
out individual cognitive functions (“weak AI”) [19].

(a)(a) (b) (c)(c)

Figure1.
Figure
Figure 1.1.Alan
AlanTuring
Alan Turingat
Turing atage
at age 16
age 16 (a).
16 (a). Source:
Source:Archive
Source: ArchiveCentre,
Archive Centre,King’s
Centre, King’s
King’sCollege,
College,
College,Cambridge.
Cambridge.
Cambridge. TheThe
Papers
The of of
Papers
Papers of
AlanTuring,
Alan Turing,AMT/K/7/4.
AMT/K/7/4. The
The same image
image after
afterapplying
applyinga aSobel
Sobelfilter inin
filter thethe
x (b)
x and
(b) y (c)
and y direction.
(c) direction.
Alan Turing, AMT/K/7/4. The same image after applying a Sobel filter in the x (b) and y (c) direction.

Figure2.2.Timeline
Figure Timelineof
ofAI
AI (orange) and of
(orange) and of AI
AIin
inmedicine
medicine(blue).
(blue).

Figure 2. Timeline of AI (orange) and of AI in medicine (blue).

In a public lecture held in 1947, Turing first mentioned the concept of a “machine that
can learn from experience” [12] and posed the question “can machine think?” in his seminal
In aentitled
paper public“Computing
lecture held machinery
in 1947, Turing first mentioned
and intelligence” [13].the concept
In this of the
paper, a “machine
“imitationthat
can learnalso
game”, from experience”
referred [12]
to as the and posed
Turing theproposed
test, was question to
“can machine
define think?”can
if a machine in his seminal
think. In
paper entitled “Computing machinery and intelligence” [13]. In this paper, the
this test, a human interrogates another human and the machine alternatively. If it is not “imitation
Cancers 2024, 16, 3702 4 of 23

2.2. Neural Networks

Neural networks were first conceived in 1943 by Warren S. McCullough (Orange,
NJ, USA, 1898–Cambridge, MA USA, 1969), a neuroscientist, and Walter Pitts (Detroit,
USA, 1923–USA, 1969), a logician, as an abstract model to describe the functioning of
the brain [20]. It consisted of a network of units (nodes) that simulate the brain cells,
the neurons, which could receive a limited number of binary inputs and send a binary
output to the environment or other neurons [21,22]. This early model was not designed
to be able to learn [23], as the weights of its units and signals were fixed, in contrast
to modern ML networks, which have learnable weights [24]. In 1949, the psychologist
Donald O. Hebb (Chester, Canada 1904–Chester, Canada, 1985) introduced the first rule
for self-organized learning: “any two cells or systems of cells that are repeatedly active at
the same time will tend to become associated: so that activity in one facilitates activity in
the other” [25], meaning that the weights of connections are increased when frequently
used simultaneously [26]. Inspired by these works, Marvin L. Minsky and Dean Edmonds
built an analog neural network machine called “stochastic neural-analog reinforcement
calculator” (SNARC), which could determine a way out of a maze [27].
Shortly afterward, a learning neural network machine, the perceptron (Figure 3), was
developed by the psychologist Frank Rosenblatt (New Rochelle, NY, USA,
1928–Chesapeake Bay, USA, 1971) [28]. Since it was built for the classification of a bi-
nary image generated using a camera, it can be considered the first application of AI to
images [29]. It used a Heaviside step as an activation function, which converts an analog
signal into a digital output [30]. The learning was accomplished by a delta rule, where
delta is simply the difference in the network output and the true value, and an incorrect
response is used to modify the weights of the connections towards the correct pattern of
prediction [30]. If multiple units are organized in a layer to be able to produce multiple
outputs from the same input, we obtain a structure that is today called single-layer artificial
Cancers 2024, 16, x FOR PEER REVIEWneural networks (ANNs) [5]. In 1960, the adaptive linear neuron (ADALINE) used 5 of the
24
weighted sum of the inputs to adjust weights [31] so that it could also estimate how much
the answer was correct or incorrect in a classification problem [32].

Figure
Figure3.3.Scheme
Schemeofofthe
theperceptron.
perceptron.

2.3.Supervised
2.3. SupervisedandandUnsupervised
UnsupervisedML ML
ML is used to explore data (‘datamining’)
ML is used to explore data (‘data mining’)totoidentify
identify variables
variablesof interest andand
of interest uncover
un-
usefuluseful
cover correlations and patterns
correlations without
and patterns any predefined
without any predefinedhypothesis to test.
hypothesis to In this
test. Insense,
this
ML operates
sense, inversely
ML operates to traditional
inversely statistical
to traditional approaches,
statistical which
approaches, begin
which withwith
begin a hypothe-
a hy-
sis [33]. The most common approach is supervised learning, where
pothesis [33]. The most common approach is supervised learning, where the system uses the system uses training
data with
training corresponding
data groundground
with corresponding truth labels
truthto learntohow
labels learntohow
predict these labels
to predict these [34].
labelsIn
unsupervised ML, the training data have no ground truth labels, and the
[34]. In unsupervised ML, the training data have no ground truth labels, and the ML learns ML learns patterns
or relationships
patterns in the data,
or relationships resulting
in the in data-driven
data, resulting solutionssolutions
in data-driven for dimensionality reduction,
for dimensionality
data partitioning, and the detection of outliers. To the first category
reduction, data partitioning, and the detection of outliers. To the first category belongsbelongs the principal
the
component analysis, PCA [35], which uses an orthogonal linear transformation
principal component analysis, PCA [35], which uses an orthogonal linear transformation to to convert
the datathe
convert into a new
data into coordinate system system
a new coordinate to perform data dimension
to perform reduction
data dimension [36]. PCA
reduction [36].is
useful when a high number of variables may cause ML models to overfit.
PCA is useful when a high number of variables may cause ML models to overfit. Overfitting Overfitting occurs
occurs when a model memorizes the training examples but performs poorly on independ-
ent test sets due to a lack of generalization capability [34].

2.4. First Applications of AI to Medicine: Expert Systems

In 1969, Minsky and Papert proved [37] that a single-layer ANN was not able to solve
Cancers 2024, 16, 3702 5 of 23

when a model memorizes the training examples but performs poorly on independent test
sets due to a lack of generalization capability [34].

2.4. First Applications of AI to Medicine: Expert Systems

In 1969, Minsky and Papert proved [37] that a single-layer ANN was not able to
solve classification problems where the separation function is nonlinear [38]. Given their
undisputed authority in the field, the interest and funding in neural networks decreased
until the early 1980s, leading to the “first AI winter”. During this era, researchers tried to
develop systems that could operate in narrower areas. The idea initially came to Edward
A. Feigenbaum (Weehawken, NJ, USA, 1936–), who became interested in creating models
of the thinking of scientists, especially the processes of empirical induction by which
hypotheses and theories were inferred from knowledge in a specific field [39]. As a result,
he developed expert systems, computer programs that make a decision such as a medical
diagnosis using a knowledge database, and a set of IF-THEN rules [21]. The first was the
DENDRAL [40], which could derive molecular structure from mass spectrometry data by
using an extensive set of rules [27]. One of the first prototypes to demonstrate the feasibility
of applying AI to medicine was CASNET, a software to provide support on diagnosis and
treatment recommendations for glaucoma [41]. MYCIN was designed to provide disease
identification and antibiotic treatment based on an extensive set of rules and patient data.
It was superseded by EMYCIN and the more general purposing INTERNIST-1 [42]. These
systems aiming at supporting the clinician’s decision are called computer decision support
systems (CDSSs).
Expert systems were limited by poor performance in areas that cannot be easily
represented by logic rules, such as detecting objects with significant variability in images.
In addition, they cannot learn from new data and update their rules accordingly, resulting
in a lack of adaptability [10]. For these reasons, in the 1990s, the interest shifted to ML, as
the larger availability of microcomputers coincided with the development of new popular
ML algorithms such as SVMs and ensemble decision trees [43].

3. Early Applications of AI to Imaging: Classical ML and ANNs

Gwilym S. Lodwick, in 1963, calculated the probability of bone tumor diagnosis with
good accuracy based on observations such as the lesion’s location relative to the physis
and whether it was in a long or flat bone [44] using an ML algorithm, specifically the Bayes
rule [45]. This early attempt can be considered the first application of ML to medical images.
Due to its low computational cost, this approach—by extracting descriptive features from
images and then analyzing them with ML models—dominated the AI field for many years
until the advent of deep learning.

3.1. Decision Tree Learning

A decision tree is a set of rules for partitioning data according to their attributes or
features (Figure 4a). This is a very intuitive process. In fact, the first classification tree is the
Porphyrian tree, a device by the 3rd-century Greek philosopher Porphyry (Tyre, Roman
Empire, present-day Lebanon, 234–Rome, 305) to classify living beings. Decision trees can
be combined with ML in decision tree learning, where one or more decision trees are grown
to create partitions of data according to rules based on the data features for classification or
regression of data.
The first attempt at decision tree methodology can likely be traced to the mid-1950s
with the work of the statistician William A. Belson. He aimed to predict the degree of
knowledge viewers had about the “Bon Voyage” television broadcast by using demographic
variables such as occupational and educational levels [46], overcoming the limitations of
linear regression [47–49]. Later, J.N. Morgan and J.A. Sonquist [50] proposed what is
now considered the first decision tree method, popularized thanks to the AID computer
program [49]. Impurity is a measure of the class mix of a subset, and splits are chosen so
that the decrease in impurity is maximized. Currently, the Gini index [51] is the preferred
 [45]. This early attempt can be considered the first application of ML to medical images. Due
to its low computational cost, this approach—by extracting descriptive features from images
and then analyzing them with ML models—dominated the AI field for many years until
the advent of deep learning.
Cancers 2024, 16, 3702 6 of 23
3.1. Decision Tree Learning
A decision tree is a set of rules for partitioning data according to their attributes or
features
method for(Figure 4a). This
measuring is a very
impurity; intuitive process.
it represents In fact, the
the probability of afirst classification
randomly chosentree is the
element
Porphyrian
being tree, classified
incorrectly a device byso the
that3rd-century
a value of zeroGreek
meansphilosopher
a completelyPorphyry (Tyre, Roman
pure partition. The
Classification And Regression Trees−CART by Leo Breiman also used pruning, atrees
Empire, present-day Lebanon, 234–Rome, 305) to classify living beings. Decision can
process
be combined with ML in decision tree learning, where one or more decision
that reduces the tree size to avoid overfitting [52,53]. It is still largely used in imaging trees are grown
to createdue
analysis partitions of data according
to its intuitiveness and ease to of
rules
use,based
e.g., iton the data
could features
classify tumorfor classification
histology from
or regression
image of data.
descriptions in MRI [54].

(a) (b)
Figure4.4.Application
Figure Applicationofofdecision
decisiontrees
trees(a)
(a)and
andsupport
supportvector
vectormachines
machines(b)(b) learning
learning toto
thethe classifi-
classifica-
tion of iris flower species from petal width and length. Prediction (areas) and training data (dots)(dots)
cation of iris flower species from petal width and length. Prediction (areas) and training data and
and the resulting decision tree are shown on the left and right sides, respectively.
the resulting decision tree are shown on the left and right sides, respectively.

3.2. Support Vector

The first Machines
attempt and Other
at decision Traditional ML
tree methodology canApproaches
likely be traced to the mid-1950s with
the work of the
The aim ofstatistician William
Support Vector A. Belson.(SVMs)
Machines He aimed is totodetermine
predict the adegree of knowledge
hyperplane in the
viewers had about
n-dimensional theof“Bon
space Voyage”
attributes television
that separates broadcast
data into by two
usingordemographic
more classesvariables
for the
such as occupational
purpose of classification.and The
educational
searchedlevels [46], overcoming
hyperplane is such that thethe
limitations
minimumofdistance
linear regres-
from
sion [47–49]. Later, J.N. Morgan and J.A. Sonquist [50] proposed what
it to the convex hull (i.e., the minimum enclosing a set of points [55]) of classes is maximal.is now considered the
first decision tree method, popularized thanks to the AID computer
This idea was first proposed by Vladimir Vapnik and Alexey Chervonenkis in 1964 [56]. A program [49]. Impurity
is a measure
few of the
years later, class mix
Corinna of aand
Cortes subset, and splits
Vladimir Vapnik are chosen so that the
[57] proposed thedecrease in impu-
first soft-margin
rity isThe
SVM. maximized. Currently,
latter allows the Gini
the inclusion of index [51]number
a certain is the preferred methoddata
of misclassified for measuring
while keepingim-
purity;
the it represents
margin as wide as thepossible
probability
so thatof aother
randomly
pointschosen
can stillelement being incorrectly
be classified classi-
correctly. SVMs
fiedalso
can so that a value
perform of zero means
nonlinear a completely
classification usingpure partition.
the “kernel The Classification
trick”, a mapping toAnd Re-
higher
gression Trees−CART
dimensional feature spaceby Leo Breiman
using properalso used pruning,(e.g.,
transformations a process that reduces
polynomial the tree
functions, size
radial
basis functions).
to avoid SVMs
overfitting are one
[52,53]. It isof thelargely
still most frequently
used in imagingused ML approaches
analysis due to in
itsmedical data
intuitiveness
analysis
and ease[58],
of use,and they
e.g., have classify
it could been found, tumor for instance,from
histology to provide accurate results
image descriptions for[54].
in MRI the
classification of prostate cancer from multiparametric MRI [59]. SVM classifiers were also
widely used in the analysis of neuroimaging data, e.g., in the study of neurodevelopment
disorders [60] and of neurodegeneration [61]. An example of SVM classification is provided
in Figure 4b.
Naïve Bayes learning, another popular ML approach, involves constructing the proba-
bility of assigning a class to a vector of features based on Bayes’ theorem and then assigning
the class with the maximum probability [62,63]. The K-nearest neighbors (KNNs) algorithm
was introduced by T. Cover and P. Hart in 1967 [64]. Its formulation appears to have been
made by E. Fix and J.L. Hodges in a research project carried out for the United States armed
forces, which introduced discriminant analysis, a non-parametric classification method [65].
They investigated a rule that might be called the KNN rule, which assigns to an unclassified
sample point the classification of the nearest of a set of previously classified points.
Traditional machine learning (ML) models analyze input data structured as vectors of
attributes, also known as variables, descriptors, or features. These features can be either
semantic (e.g., “spiculated lesion”) or agnostic (quantitative) [66]. Coding a problem in
Cancers 2024, 16, 3702 7 of 23

terms of a feature vector can lead to an extremely large number of features depending on the
complexity of the problem to address, increasing the risk of overfitting. Feature selection is
a process to determine a subset of features such as all the features in the subset are relevant
to the target concept, and no feature is redundant [67,68]. A feature is considered redundant
when adding it on top of the others will not provide additional information; for instance,
if two features are correlated, these are redundant to each other [69]. Feature selection
may be considered an application of Ockham’s razor to ML. According to Ockham’s razor
principle, attributed to the 14th-century English logician William of Ockham (Ockham,
England, 1285–Munich, Bavaria, 1347), given two hypotheses consistent with the observed
data, the simpler one (i.e., the ML model using the lower number of features), should
be preferred [70]. Depending on the type of data, feature selection can be classified as
supervised, semi-supervised, or unsupervised [69]. There are three main classes of feature
selection methods: (i) embedded feature selection, where ML includes the choice of the
optimal subset; (ii) filtering, where features are discarded or passed to the learning phase
according to their relevance; and (iii) wrapping, which requires evaluating the accuracy of
a specific ML model on different feature subsets for choice of the optimal one [67]. “Tuning”
is the task of finding optimal hyperparameters for a learning algorithm for a considered
dataset. For instance, decision trees have several hyperparameters that may influence their
performance, such as the maximum depth of the tree and the minimum number of samples
at a leaf node [71]. Early attempts for parameter optimization include the introduction of
the Akaike information criteria for model selection [72]. More recent strategies include grid
Cancers 2024, 16, x FOR PEER REVIEW 8 of 24
search, in which all parameter space is discretized and searched, and random search [73], in
which values are drawn randomly from a specified hyperparameter space, which is more
efficient, especially for ANNs [71].
3.3. First Uses of Neural Networks for Image Recognition
3.3. First Uses of Neural Networks for Image Recognition
To address the criticism of M. Minsky and S. Papert [37] and enable neural networks
To address
to solve the criticism
nonlinearly separableofproblems,
M. Minskymanyand S. Papert [37]
additional andofenable
layers neural units
neuron-like networks
must
to
besolve nonlinearly
placed between separable
input andproblems, manyleading
output layers, additional layers of neuron-like
to multilayer ANNs. Theunits
first must
work
be placed between
proposing input
multilayer and output
perceptrons layers,
was leading
published to multilayer
in 1965 ANNs.and
by Ivakhnenko TheLapa
first[71].
workA
proposing multilayer perceptrons was published in 1965 by Ivakhnenko
multilayered neural network was proposed in 1980 by Fukushima called “Neocognitron” and Lapa [71].
A[74],
multilayered neural network was proposed in 1980 by Fukushima called “Neocogni-
which was used for image recognition [75], and included multiple convolutional layers
tron” [74], image
to extract which features
was usedoffor image recognition
increasing complexity.[75], andintermediate
These included multiple
layersconvolutional
are called hid-
layers to extract image features of increasing complexity. These intermediate
den layers [21] and multilayer architectures of neural networks are called “deep”. layers are
Hence,
called hidden layers [21] and multilayer architectures of neural networks are called
the term “deep learning” (DL) was coined by R. Dechter [76]. The difference between single- “deep”.
Hence, the multilayer
layer and learning”
term “deep ANNs (DL) was
is shown in coined
Figure by
5. R. Dechter [76]. The difference between
single-layer and multilayer ANNs is shown in Figure 5.

(a) (b)
Figure5.5.Comparison
Figure Comparisonbetween
betweensingle-layer
single-layer(a)
(a)and
andmultilayer
multilayerANNs
ANNs(b).
(b).

InaaDL
In DLneural
neural network,
network, training
training isis performed
performedby byupdating
updatingall allthe
theweights
weightssimultane-
simulta-
ously in in
neously thethe
opposite
oppositedirection to a to
direction vector that indicates
a vector by theby
that indicates change in the error
the change in theif error
weights
if
are changed
weights by a small
are changed by amount to searchtoa search
a small amount minimum in the error
a minimum in response.
in the This method
error in response. Thisis
called “standard
method is called gradient
“standarddescent”
gradient[77], and one[77],
descent” of itsand
limitations
one of was that it made
its limitations tasks
was such
that it
as image recognition too computationally expensive.
made tasks such as image recognition too computationally expensive.
The invention of the backpropagation learning algorithm in the mid-1980s [78] improved
significantly the efficiency of training of neural networks. The back-propagation equations
provide us with a way of computing the gradient of the cost function starting from the final
layer [79]. Then, the backpropagation equation can be applied repeatedly to propagate gradi-
ents through all modules all the way to the input layer by using the chain rule of derivatives
to estimate how the cost varies with earlier weights [80]. This learning rule and its variants
Cancers 2024, 16, 3702 8 of 23

The invention of the backpropagation learning algorithm in the mid-1980s [78] improved
significantly the efficiency of training of neural networks. The back-propagation equations
provide us with a way of computing the gradient of the cost function starting from the
final layer [79]. Then, the backpropagation equation can be applied repeatedly to propagate
gradients through all modules all the way to the input layer by using the chain rule of
derivatives to estimate how the cost varies with earlier weights [80]. This learning rule and its
variants enabled the use of neural networks in many hard medical diagnostic tasks [81].
The introduction of the rectifier function or ReLu (rectified linear unit), an activation
function, which is zero if the input is lower than the threshold and is equal to the input
otherwise, helped reduce the risk of vanishing/exploding gradients [82] and is the most
used activation function as of today.
Despite this progress, AI entered its second winter at the beginning of the 1990s„ which
involved the use of neural networks, to the point that, at a prominent conference, it was
noted that the term “neural networks” in a manuscript title was negatively correlated with
acceptance [83]. This new AI winter was partly due to the vanishing/exploding gradient
problem of DL, which is the exponential increase or decrease in the backpropagated gradient
of the weights in a deep neural network.

3.4. Ensemble Machine Learning

During the second winter of AI in the 1990s, while neural networks saw a decrease
in interest, AI research was focused on other ML techniques, such as SVM and decision
trees, and on ways to improve their accuracy. Significant improvements in the accuracy of
decision trees arise from growing an ensemble of trees and subsequently aggregating their
predictions [84]. In bagging aggregation, to grow each tree, a random selection is made
from the examples in the training set [85], using a resampling technique approach called
“bootstrap” [86]. The bagging aggregation averages over the versions when predicting
a numerical outcome and does a plurality vote when predicting a class [87]. Random
forest, proposed in 1995 by Tin K. Ho [88], is an example of this approach. In boosting
aggregation [89], the distribution of examples is filtered in such a way as to force the weak
learning algorithm to focus on the harder-to-learn parts of the distribution. The popular
Adaboost (from ‘adaptive boosting’) reweights individual observations in subsequent
samples and is well suited for imbalanced datasets [90].

3.5. ML Applications to Medical Imaging: CAD and Radiomics

Computer-aided detection or diagnosis (CAD) systems assist clinicians by analyzing
medical images and highlighting potential lesions or suggesting diagnoses [91,92]. Early
CAD systems used hand-crafted image features, which were introduced into rule-based
algorithms to produce an index (e.g., a probability of malignancy) to be used for diagno-
sis [93]. Features could include spiculations, roughness of margins, and perimeter-to-area
ratio for distinguishing malignant breast lesions [94] or lung disease in radiography [95].
The first commercial CAD system was the ImageChecker M1000 (R2 Technology,
Los Altos, CA, USA), which received US Food and Drug Administration−FDA approval
in 1998 [96] and provided the likelihood of malignant lesions according to the presence
of clusters of bright spots and spiculated masses, also highlighting regions at risk of
malignancy [93]. Other breast CAD systems were proposed for breast magnetic resonance
imaging: CADstream (Merge Healthcare Inc., Chicago, IL, USA) and DynaCAD for breast
(Invivo, Gainesville, FL, USA) [93]. Early CAD systems exhibited lower specificity and
positive predictive value compared to double reading by radiologists, rendering the sole
use of CAD not advisable [97].
Textural features, first introduced by Robert.M. Haralick and coworkers in 1973 [98],
began to be used for the quantitative analysis of texture with ML for pattern recognition
on computed tomography [99]. The term “radiomics” first appeared in 2010 [100,101],
combining the terms “radio”, referring to radiological sciences, and the suffix “omics”,
often used in biology (e.g., genomics, transcriptomics, and proteomics), to emphasize a
Cancers 2024, 16, 3702 9 of 23

research field encompassing the entire view of a system by mining a large amount of
data [102]. Radiomics focused on investigating the tumor phenotype in imaging for build-
ing prognostic and predictive models [103], in particular for oncological applications [14].
Thus, it has largely contributed to the idea that ML can be applied to quantitatively analyze
images [104]. An array of ML techniques is currently used for radiomics, including SVM
and ensemble decision trees [1]. The radiomic approach, complemented by ML, has been
largely implemented in a large variety of studies devoted to the identification of imaging-
based biomarkers of disease severity assessment or staging and patient’s outcome or risk
for side effects [105–109]. The scientific community is still investigating the robustness and
reproducibility of radiomics features and their dependence on image acquisition systems
and parameters across different modalities [110,111].

4. The Era of Deep Learning in Medical Imaging

In 1989, Yann LeCun introduced the concept of a convolutional neural network (CNN)
to recognize handwritten digits, paving the way for the use of deep neural networks in
imaging [112]. CNNs have layers that perform a convolution operation with a kernel that
acts as a filter, e.g., a Sobel filter, whose effect is shown in Figure 1b,c. The numerical values
of the kernels that operate image filtering are not fixed a-priory; they are learned from data
and set during the training phase [93]. Unlike the traditional ML approach, the DL does
not require the extraction of meaningful features to describe the image. The performance
of CNNs can be boosted by artificially augmenting the dataset by affine transformations
of the images, like translation, scaling, squeezing, and shearing, a strategy that reduces
overfitting [83]. Further improvements were achieved with the introduction of specialized
layers, i.e., stacks of neural units that perform a particular task within the DL architecture
like dropout [113], pooling, and fully connected layers [93], allowing an endless spectrum
of configurations for the most diverse tasks. In 2012, a deep CNN developed at the Univer-
sity of Toronto demonstrated excellent performance at the ImageNet Large Scale Visual
Recognition Challenge in 1.2 million high-resolution images of 1000 classes [114]. Compe-
titions or challenges have become a major driver of progress in AI [4,115]. Autoencoders
and Convolutional Autoencoders [116] are used for dimensionality reduction, data and
image denoising, and uncovering hidden patterns in unlabeled data. Additionally, sparse
autoencoders can generate extra useful features [117].
Consequently, it is now clear that DL can be designed for almost any many domains
of science, business, and government to perform as diverse as image, signal, and sound
recognition, transformation, or production.

4.1. Medical Images Classification with Deep Learning Models

Neural networks began to be used in CAD by M.L. Giger and coworkers in ra-
diographic images [91,118,119], e.g., in lung [120] and breast [92] investigations. In the
1990s [118] researchers started to use CNNs to identify lung nodules [121,122], and detect
micro-calcifications in mammography [123]. Since the first attempts, CNNs have demon-
strated a great potential to solve a large variety of classification tasks; thus, they have
been implemented across many pathologies and imaging modalities [124]. The CardioAI
CAD system was one of the first neural network-based commercial systems for analyzing
cardiac magnetic resonance images [125]. Moreover, AI can integrate data from different
modalities, an approach termed multimodal AI or multimodal data fusion, which can, for
instance, be used to diagnose cancer using both imaging and patient EHR data. This task
can be accomplished by designing neural networks that accept multiple data, resulting in
multidimensional analysis [126]. Recently, the You Only Look Once (YOLO) neural network
architecture, initially designed for real-time object detection, has been under investigation
for polyp detection in colonoscopy [127] and identifying dermoscopic and cardiovascular
anomalies [128].
Cancers 2024, 16, 3702 10 of 23

Among the CNN-based systems approved for clinical use, ENDOANGEL (Wuhan
EndoAngel Medical Technology Company, Wuhan, China) can provide an objective assess-
ment of bowel preparation every 30 s during the withdrawal phase of a colonoscopy [129].
The CNN-based system can also analyze images to predict overall survival and occurrence
of distant metastases [130].
These DL models are characterized by a very large number of free parameters that
must be set during the training phase, making network training from scratch computa-
tionally intensive. In transfer learning [131,132], the knowledge acquired in one domain
is transferred to a different one, much like a person using their guitar-playing skills to
learn the piano [133]. This allows a learner in one domain (e.g., radiographs) to leverage
information previously acquired by models such as the Visual Geometry Group (VGG)
and Residual Network (ResNet), which were trained on a related domain (e.g., images of
common objects).

4.2. Segmentation with Deep Learning Models

AI can be used in medical image analysis to automatically subdivide an image into
several regions based on the similarity or difference between regions, thus performing
segmentation of soft tissues and lesions, a tedious task if performed manually. Initially,
segmentation was pursued using semiautomated approaches, such as edge-, region-, or
threshold-based segmentation. For instance, Sobel filters were applied to enhance and
detect lesion borders. However, these methods are highly sensitive to noise and image
contrast, limiting their effectiveness [134]. Then, automated segmentation was attempted by
using unsupervised [135] or, supervised ML [136]. Another approach consists of calculating
radiomic features in the neighborhood of pixels and then classifying them using ML [137].
Image segmentation was made more efficient by the U-Net CNN [138] currently used for a
large variety of segmentation tasks across different image modalities [139–144]. It consists
of an encoder branch where the input layer is followed by several convolutional and
pooling layers, as in a CNN architecture for image classification; then, a symmetric decoder
branch allows obtaining a segmentation mask with the same dimension of the input image.
The U-Net’s peculiar skip connections bridge the encoder and decoder, directly transferring
detailed spatial information to the upsampling path enabling precise object locations in the
final segmentation masks [145]. In 2016, Ö. Çiçek and coworkers [146] presented a modified
three-dimensional version of the original U-Net (3D U-Net) for volumetric segmentation.

4.3. Medical Image Synthesis: Generative Models

Generative Adversarial Networks (GANs) represent a DL architecture capable of
generating new and realistic images by training from a dataset of images, video, or other
types of data [147]. GANs include two DL networks, a generator, and a discriminator that
are trained in an adversarial way, the target goal being to train the generator to produce an
image realistic enough to induce the discriminator into classifying it as real. By using GANs
it is possible to generate images from other images or from a text string, for instance. One
of the most recent architectures for image generation is stable diffusion, which can produce
high-quality images from text or text-conditional images, e.g., “basal cell carcinoma” in a
dermoscopic image [148].
The use of generative models has proven to be valuable in medical imaging appli-
cations, including data synthesis or augmentation [149,150]. An example could be the
generation of pseudo-healthy images, images that visualize a negative image of a patient
that is being examined in order to facilitate lesion detection [151]. Virtual patient cohorts
can be synthesized to perform virtual clinical trials for testing test new drugs, therapies, or
diagnostic interventions, thus reducing the cost of clinical trials on humans [152].
Other applications include image denoising and artifact removal [153], image trans-
lation between different modalities [154], and multi-site data harmonization [155]. Fast
AI-based image reconstruction also emerged to allow real-time MRI imaging [156]. GAN-
based MRI image reconstruction particularly excels in capturing fine textures [157]. CNNs
Cancers 2024, 16, 3702 11 of 23

have also been applied to rigid [158] and deformable image registration [159], which is
necessary to precisely track the absorbed dose in radiotherapy treatments at the voxel
level [160]. A new promising neural architecture, the neural fields, can perform efficiently
any of the above tasks by parameterizing the physical properties of images [161].

4.4. From Natural Language Processing to Large Language Models

Recurrent neural networks (RNNs) process an input sequence one element at a time,
maintaining in their hidden units a ‘state vector’ that implicitly contains information
about the history of all the past elements of the sequence. RNNs can predict the next
character in a text or the next word in a sequence [162], making them useful for speech
and language tasks. However, their training is challenging because the backpropagated
gradients can either grow or shrink at each time step. Over many time steps, this can lead
to gradients that either explode or vanish [163]. In 1997, long short-term memory (LSTM)
RNNs were invented [164], which solve the problem of vanishing gradients for sequences
of symbols by including a forget gate, which allows the LSTM to reset its state [165,166].
This subfield of AI aiming at developing the computer’s abilities to understand or generate
human language is called natural language processing (NLP) [167]. There has been a surge
of research in NLP diagnostic models from structured or unstructured electronic health
records (EHR) [168–170]. In 2007, IBM introduced Watson, a powerful NLP software which,
in 2017, was instrumental in identifying new RNA-binding proteins linked to amyotrophic
lateral sclerosis [125].
A breakthrough in this field was the Transformer DL architecture, which, by em-
ploying the self-attention mechanism [171] showed excellent capabilities in managing
dependencies between distant elements in an input sequence and in exploiting parallel
processing to reduce execution times. Transformers are the basic components of Large
Language Models (LLM), such as the Generative Pretrained Transformers (GPT) by Ope-
nAI or the Bidirectional Encoder Representations from Transformers (BERT) by Google.
These are trained on a large amount of data from the web and are able to generate text
to make translation, summarization, and complete sentences, and also the production of
creative content in domains specified by the users. LLM can be used as a decision support
system that recommends appropriate imaging from a patient’s symptoms and history [172].
Recently, GPT4, a new version of the ChatGPT by OpenAI, a generative LLM that can
generate human-like answers, was released. GPT4 can analyze images, implying that, if
successfully applied to radiology, it could writes diagnoses from images [172] and can act
as a virtual assistant to the radiologist [173].
A transformer-based encoder-decoder model analyzing chest radiograph images to
produce radiology report text was assessed by comparing its generated reports with those
generated by radiologists [174].
Beyond the automated image interpretation tasks, since the public availability of the
ChatGPT chatbot at the end of 2022, its potential use, for example, in assisting clinicians in
the generation of context-aware descriptions for reporting tasks, became apparent [172].
Similar uses entail a series of implications that are much debated in the community [175].
Moreover, LLMs can aid in patients comprehending their reports by summarizing informa-
tion at any reading level and in the patient’s preferred language [173]. The architectures
initially developed to understand and generate text have soon been adapted for other
tasks in several domains, including computer vision. Vision Transformers (ViTs) [176] are
a variant of transformers specifically designed for computer vision tasks, such as image
classification, object detection, and image generation. Instead of processing sequences of
word tokens (i.e., elements of textual data such as words and punctuation marks) as they
do in NLP tasks, ViTs process image patches to accomplish relevant tasks in medical image
analysis, such as lesion detection, image segmentation, registration, and classification [171].
The image generation processes operated by GANs could be further enhanced by imple-
menting the attention mechanisms [167]. Combining ViTs for analyzing diagnostic images
with LLMs for clinical report interpretation could result in a comprehensive image-based
Cancers 2024, 16, 3702 12 of 23

decision support tool. An extremely appealing use of Transformers is their potential to

handle multimodal input data. This capability was demonstrated in the work by Akbari
et al. [177], where a transformer-based architecture, the Video-Audio-Text Transformer, was
developed to integrate images, audio, and text.
The possibility of implementing modality-specific embedding to convert the entries
of each modality to processable information for a transformer-like architecture opens the
possibility of analyzing in a single framework heterogeneous data types, which would be
extremely relevant for medical applications where complementary information is encoded in
textual clinical reports and tests, medical imaging, genetic and phenotypic information [152].

4.5. Foundational Models

A limitation of the AI-based tools discussed so far is their ‘narrow scope’, as they
are typically designed to detect specific image abnormalities [178]. In contrast, models
like GPT-4 are pre-trained on vast and diverse datasets that encompass text, audio, and
images, allowing for broader applications. These are referred to as ‘foundation models’
because they can be fine-tuned for specific tasks using transfer learning, serving as the
foundation for models capable of addressing specialized tasks [179]. This is a radical shift
from previous artificial intelligence tools that were designed to solve specific tasks [180].
Recently, a foundation model trained on ImageNet, a large database of natural images
(https://www.image-net.org/), was fine-tuned to generate realistic chest x-ray images
based on prompts from the user [181]. A foundation model, after fine-tuning for pathology,
was capable of nuclear segmentation, primary and metastatic cancer detection, cancer
grading, and sub-typing, outperforming previous state-of-the-art models [182].
Since foundation models can perform various tasks across diverse domains, they can
be adapted through in-context learning—introduced in 2020 with the GPT-3 language
model—where the model learns from user-provided text explanations (or ‘prompts’) with
a few examples [180]. In this way, models can adapt to new distributions of data on the
fly using limited data, whereas traditional AI models need extensive retraining on a new
dataset. A hospital, for instance, can teach a model to interpret X-rays from a brand-new
scanner simply by providing prompts that show a small set of examples [180].
GPT-4 was instructed by users who constructed textual prompts from 25 CT radiology
reports. The GPT-4 was then able to perform various tasks on unseen reports, such as
extracting lesion parameters and identifying metastatic disease with high accuracy [183].
In this way, foundation models can circumvent the problem of data scarcity in medical
imaging [184]. Another mechanism for this purpose is self-supervised learning, where the
models build data representations by solving pretext tasks. Pretext tasks are tasks whose
outcome is not of interest, such as image colorization, but result in the model learning
representations of input images, improving its generalizability [185].

5. Open Challenges and Pathways for AI in Medical Imaging

In 2017, for the first time in history, DeepMind’s AlphaGo, a self-trained system
based on a deep neural network, beat the world champion in arguably the most complex
board game (called “Go”), thus achieving superhuman performance [186]. It is no wonder,
then, that AI has achieved physician-level accuracy in a broad variety of diagnostic tasks,
including image recognition [2], segmentation, and generation [187,188]. The development
of graphics processing units (GPUs) and cloud computing has provided the significant
computational power required to train DL models on large datasets. Additionally, they
have made it possible to perform AI-driven tasks in real-time, such as image registration
and reconstruction for tumor tracking in radiotherapy [189].
Alongside neural networks, other ML techniques remain largely used due to their
ability to accurately solve classification and regression problems without the need for
expensive computational resources [1]. ML can also assist in diagnosing conditions from
clinical data, such as myocardial infarction [190,191] and in making differential diagnoses
among various findings in neonatal radiographs [192]. For instance, decision trees have
Cancers 2024, 16, 3702 13 of 23

demonstrated the ability to predict specific phenotypes from raw genomic data [193], to
assign emergency codes based on symptoms during triage in emergency departments [194],
and other tasks [195].
Since recently, multi-input AI models can merge and mine the complementary infor-
mation encoded in omics data, EHRs, imaging data, phenomics, and environmental data of
the patient, which represent a current technical challenge [152]. Meta AI’s Imagebind [196]
and Google Deepmind’s Perceiver IO [197] represent significant advancements in pro-
cessing and integrating multimodal data. Other than the flexibility of architecture, other
reasons contributed to the large adoption of DL in medical imaging [198–200]. Despite
these successes, there are also challenges. This section explores both the factors contributing
to AI’s success and the concerns surrounding its use.

5.1. Open-Source Libraries and Databases

The availability of open-source and free libraries has greatly facilitated the adoption of
DL by researchers and clinicians. TensorFlow (https://www.tensorflow.org) and PyTorch
(https://pytorch.org), both of which can be run using Python (www.python.org), a high-
level and easy-to-learn language, are widely used for developing DL-based segmentation
and classification tools. This trend has been highlighted in numerous systematic surveys [1].
Medical Open Network for Artificial Intelligence (MONAI) [201,202] is a project
initiated in 2020 by NVIDIA and King’s College London, which has since evolved into
the MONAI framework. This open-source framework, built on PyTorch, focuses on DL in
healthcare imaging. Its goal is to develop and share best practices for AI in medical imaging,
offering state-of-the-art, end-to-end training workflows. MONAI provides researchers
with an optimized and standardized approach to creating and evaluating DL models.
The framework includes workflows for utilizing domain-specific networks, loss functions,
metrics, and optimizers [203].
The research community increasingly tends to share their programming code, often
organized as easy-to-run scripts with clear instructions, alongside their datasets. This
practice aims to make research studies more reproducible [204,205]. Sharing data, including
raw and processed images (with segmentations and annotations) and clinical data, in public
repositories is also highly encouraged [206–209]. Public medical databases, like the Cancer
Imaging Archive [210], can be used to train and validate new DL models.

5.2. Real World Evaluation

A significant portion of AI tools lacks evidence of efficacy published in peer-reviewed
scientific journals [211]. Additionally, the performance achieved during the research phase
is often difficult to replicate in clinical settings [212].
In a review of 2021, out of the AI-powered medical devices approved or cleared by
the US Food and Drug Administration, only a small number have been tested through
prospective randomized controlled trials [213]. Therefore, for AI tools to be integrated into
clinical practice, systematic clinical evaluations or trials are necessary.
Many clinical trials were introduced at the end of 2023, as pointed out by a recent
review: eighty-six randomized clinical trials for AI-based tools were registered, mostly di-
agnostic, predominantly as single-center trials [214,215]. The ScreenTrustCAD prospective
clinical trial demonstrated that AI alone was non-inferior to double reading by two radiol-
ogists in screening mammography. Additionally, combining AI with a single radiologist
outperformed double reading by two radiologists, likely due to AI’s high sensitivity in
detecting cancer and the ability of consensus readers to improve specificity by dismissing
AI-generated false positives [2].

5.3. Explainability/Interpretability
A key barrier to the widespread adoption of AI-based tools in medical imaging is that
these systems are often viewed as black boxes, making it challenging to understand how
they arrive at their decisions [216].
Cancers 2024, 16, 3702 14 of 23

Interpretability and explainability in AI relate to understanding how an AI system ar-

rives at its decisions or outputs [217]. Although these terms are often used interchangeably,
interpretability generally refers to either designing models that inherently reveal insights
into the patterns they learn during training or analyzing the model to uncover relationships
it has identified, such as by examining feature importance [218]. Techniques to enhance
interpretability include visual methods like saliency maps and heatmaps, which highlight
areas an image-based model considers significant for its predictions [219]. Explainability,
in contrast, focuses on making the AI’s decision-making process more understandable and
communicable to humans [216,220].
The scientific community is developing explainability frameworks to make AI models
more transparent and understandable to humans, leading to “explainable AI” (XAI). Both
system developers and end users can obtain, in this way, an idea of the motivations behind
the decision provided by an AI system. XAI is essential in DL applications to medical
imaging to ensure transparency, accountability, safety, regulatory compliance, and clinical
applicability [216].
By providing interpretable explanations for DL model predictions, XAI techniques
enhance the trust, acceptance, and effectiveness of AI systems in healthcare, ultimately
improving patient outcomes and advancing the field of medical imaging.

5.4. Ethical Issues

AI also brings risks and ethical issues, such as the need to ensure fairness, meaning it
should not be biased against some group or minority [221]. Bias in AI software may result
from unbalanced training data. For instance, due to a severe imbalance in the training data,
an AI tool for diagnosing diabetic retinopathy was found to be more accurate for light-
skinned subjects than for dark-skinned subjects [222]. Another risk arises from differences
between the training data used to develop the algorithm and actual patient data, known
as data shift. Changes in patient or disease characteristics or technical parameters (e.g.,
treatment management and imaging acquisition protocol) over time or across locations
can affect the accuracy of AI (input data shift) [223]. Additional risks related to AI include
cybersecurity challenges [224]. Implementing AI requires managing these risks through a
quality assurance program and quality management system [225]. Within the European
Union (EU), the regulatory framework for medical devices is defined by the European
Medical Devices Regulation (EU) 2017/745 (EU MDR) and the General Data Protection
Regulation (EU) 2016/679 (GDPR), which established criteria for AI implementation [3,226].
Furthermore, the EU has proposed legislation known as “The Artificial Intelligence Act (AI
Act)” [227], aiming at creating a unified regulatory and legal structure for AI.

6. Conclusions
After a long and tumultuous history, we are in a phase of enthusiasm and promises
regarding AI applications to medicine. Fueled by its versatility, impressive results, and
the availability of powerful computing resources and open-source libraries, AI is one of
the most promising frontiers in medicine. Some medical imaging tasks can be successfully
addressed by traditional ML methods like RF, which is less prone to overfitting than DL and
more easily interpretable. Various DL architectures can efficiently and accurately perform a
range of tasks, including image reconstruction and registration. DL networks have also
achieved human-level performance in tasks such as lesion detection, image classification,
and segmentation. Additionally, foundation models, pre-trained on a large scale, can be
fine-tuned for diverse domains, requiring less training data than training a DL model
from scratch.
Indeed, to facilitate the diffusion of AI-based tools in clinical workflows, in addition to
the development of increasingly cutting-edge technological solutions that can answer dif-
ferent clinical questions, AI-based systems should be validated in large-scale clinical trials
to demonstrate their effectiveness. Additional concerns regarding AI in healthcare must be
addressed, including the view of AI tools as ‘black boxes’, which calls for more interpretable
Cancers 2024, 16, 3702 15 of 23

and explainable models to earn the trust of both doctors and patients. Ethical issues, such
as ensuring fairness and reliability in AI systems, also need careful consideration.

Author Contributions: Conceptualization, M.A.; methodology, A.R.; writing—original draft prepa-

ration, M.A. and J.S.; writing—review and editing, A.R.; visualization, G.P.; supervision, A.D. All
authors have read and agreed to the published version of the manuscript.
Funding: This work was supported by the Italian Ministry of Health (Ricerca Corrente 2024)
[J33C23003340001] and PNRR-M4C2-Inv. 1.3, PE00000013-“FAIR-Future Artificial Intelligence Research”-
Spoke 8 “Pervasive AI”.
Conflicts of Interest: Author Joseph Stancanello is employed by the company Elekta SA. The
remaining authors declare that the research was conducted in the absence of any commercial or
financial relationships that could be construed as a potential conflict of interest.

References
1. Avanzo, M.; Porzio, M.; Lorenzon, L.; Milan, L.; Sghedoni, R.; Russo, G.; Massafra, R.; Fanizzi, A.; Barucci, A.; Ardu, V.; et al.
Artificial Intelligence Applications in Medical Imaging: A Review of the Medical Physics Research in Italy. Phys. Med. 2021, 83,
221–241. [CrossRef] [PubMed]
2. Dembrower, K.; Crippa, A.; Colón, E.; Eklund, M.; Strand, F. Artificial Intelligence for Breast Cancer Detection in Screening
Mammography in Sweden: A Prospective, Population-Based, Paired-Reader, Non-Inferiority Study. Lancet Digit. Health 2023, 5,
e703–e711. [CrossRef] [PubMed]
3. Zanca, F.; Brusasco, C.; Pesapane, F.; Kwade, Z.; Beckers, R.; Avanzo, M. Regulatory Aspects of the Use of Artificial Intelligence
Medical Software. Semin. Radiat. Oncol. 2022, 32, 432–441. [CrossRef] [PubMed]
4. Armato, S.G.; Drukker, K.; Hadjiiski, L. AI in Medical Imaging Grand Challenges: Translation from Competition to Research
Benefit and Patient Care. Br. J. Radiol. 2023, 96, 20221152. [CrossRef] [PubMed]
5. Radanliev, P.; de Roure, D. Review of Algorithms for Artificial Intelligence on Low Memory Devices. IEEE Access 2021, 9,
109986–109993. [CrossRef]
6. Samuel, A. Some Studies in Machine Learning Using the Game of Checkers. IBM J. Res. Dev. 1959, 3, 210–229. [CrossRef]
7. Hassabis, D. Artificial Intelligence: Chess Match of the Century. Nature 2017, 544, 413–414. [CrossRef]
8. Fron, C.; Korn, O. A Short History of the Perception of Robots and Automata from Antiquity to Modern Times. In Social Robots:
Technological, Societal and Ethical Aspects of Human-Robot Interaction; Korn, O., Ed.; Springer International Publishing: Cham,
Switzerland, 2019; pp. 1–12; ISBN 978-3-030-17107-0.
9. Common Sense, the Turing Test, and the Quest for Real AI. The MIT Press. Available online: https://mitpress.mit.edu/books/
common-sense-turing-test-and-quest-real-ai (accessed on 7 July 2022).
10. Haenlein, M.; Kaplan, A. A Brief History of Artificial Intelligence: On the Past, Present, and Future of Artificial Intelligence. Calif.
Manag. Rev. 2019, 61, 000812561986492. [CrossRef]
11. Muggleton, S. Alan Turing and the Development of Artificial Intelligence. AI Commun. 2014, 27, 3–10. [CrossRef]
12. Turing, A. Lecture on the Automatic Computing Engine (1947); Oxford University Press: Oxford, UK, 2004.
13. Turing, A.M. I.—Computing Machinery and Intelligence. Mind 1950, LIX, 433–460. [CrossRef]
14. Reaching for Artificial Intelligence: A Personal Memoir on Learning Machines and Machine Learning Pioneers—UNESCO Digital
Library. Available online: https://unesdoc.unesco.org/ark:/48223/pf0000367501?posInSet=38&queryId=2deebfb4-e631-4723-
a7fb-82590e5c3eb8 (accessed on 24 August 2022).
15. Understanding AI—Part 3: Methods of Symbolic AI. Available online: https://divis.io/en/2019/04/understanding-ai-part-3-
methods-of-symbolic-ai/ (accessed on 20 July 2022).
16. Garnelo, M.; Shanahan, M. Reconciling Deep Learning with Symbolic Artificial Intelligence: Representing Objects and Relations.
Curr. Opin. Behav. Sci. 2019, 29, 17–23. [CrossRef]
17. Sorantin, E.; Grasser, M.G.; Hemmelmayr, A.; Tschauner, S.; Hrzic, F.; Weiss, V.; Lacekova, J.; Holzinger, A. The Augmented
Radiologist: Artificial Intelligence in the Practice of Radiology. Pediatr. Radiol. 2021, 52, 2074–2086. [CrossRef] [PubMed]
18. McCarthy, J.; Minsky, M.L.; Rochester, N.; Shannon, C.E. A Proposal for the Dartmouth Summer Research Project on Artificial
Intelligence, August 31, 1955. AI Mag. 2006, 27, 12. [CrossRef]
19. Wang, P. On Defining Artificial Intelligence. J. Artif. Gen. Intell. 2019, 10, 1–37. [CrossRef]
20. McCulloch, W.S.; Pitts, W. A Logical Calculus of the Ideas Immanent in Nervous Activity. Bull. Math. Biophys. 1943, 5, 115–133.
[CrossRef]
21. Basheer, I.A.; Hajmeer, M. Artificial Neural Networks: Fundamentals, Computing, Design, and Application. J. Microbiol. Methods
2000, 43, 3–31. [CrossRef]
22. Piccinini, G. The First Computational Theory of Mind and Brain: A Close Look at Mcculloch and Pitts’s “Logical Calculus of
Ideas Immanent in Nervous Activity”. Synthese 2004, 141, 175–215. [CrossRef]
23. Schmidhuber, J. Deep Learning in Neural Networks: An Overview. Neural Netw. 2015, 61, 85–117. [CrossRef]
24. Wang, H.; Raj, B. On the Origin of Deep Learning. arXiv 2017, arXiv:1702.07800. [CrossRef]
Cancers 2024, 16, 3702 16 of 23

25. Hebb, D.O. Organization of Behavior. New York: Wiley, 1949, Pp. 335, $4.00. J. Clin. Psychol. 1950, 6, 307. [CrossRef]
26. Chakraverty, S.; Sahoo, D.M.; Mahato, N.R. Hebbian Learning Rule. In Concepts of Soft Computing: Fuzzy and ANN with
Programming; Chakraverty, S., Sahoo, D.M., Mahato, N.R., Eds.; Springer: Singapore, 2019; pp. 175–182. ISBN 9789811374302.
27. Toosi, A.; Bottino, A.G.; Saboury, B.; Siegel, E.; Rahmim, A. A Brief History of AI: How to Prevent Another Winter (A Critical
Review). PET Clin. 2021, 16, 449–469. [CrossRef] [PubMed]
28. Rosenblatt, F. The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain. Psychol. Rev. 1958, 65,
386–408. [CrossRef] [PubMed]
29. Parhi, K.; Unnikrishnan, N. Brain-Inspired Computing: Models and Architectures. IEEE Open J. Circuits Syst. 2020, 1, 185–204.
[CrossRef]
30. Dawson, M.R.W. Connectionism and Classical Conditioning. Comp. Cogn. Behav. Rev. 2008, 3, 115–133. [CrossRef]
31. Macukow, B. Neural Networks—State of Art, Brief History, Basic Models and Architecture. In Proceedings of the Computer Informa-
tion Systems and Industrial Management; Saeed, K., Homenda, W., Eds.; Springer International Publishing: Cham, Switzerland,
2016; pp. 3–14.
32. Raschka, S. What Is the Difference Between a Perceptron, Adaline, and Neural Network Model? Available online: https:
//sebastianraschka.com/faq/docs/diff-perceptron-adaline-neuralnet.html (accessed on 8 July 2022).
33. Butterworth, M. The ICO and Artificial Intelligence: The Role of Fairness in the GDPR Framework. Comput. Law Secur. Rev. 2018,
34, 257–268. [CrossRef]
34. Avanzo, M.; Wei, L.; Stancanello, J.; Vallières, M.; Rao, A.; Morin, O.; Mattonen, S.A.; El Naqa, I. Machine and Deep Learning
Methods for Radiomics. Med. Phys. 2020, 47, e185–e202. [CrossRef]
35. Wold, S.; Esbensen, K.; Geladi, P. Principal Component Analysis. Proc. Multivar. Stat. Workshop Geol. Geochem. 1987, 2, 37–52.
[CrossRef]
36. Dayan, P.; Sahani, M.; Deback, G. Unsupervised Learning. In Proceedings of the MIT Encyclopedia of the Cognitive Sciences; The MIT
Press: Cambridge, MA, USA, 1999.
37. Minsky, M.; Papert, S. Perceptrons; an Introduction to Computational Geometry; MIT Press: Cambridge, MA, USA, 1969; ISBN
978-0-262-13043-1.
38. Spears, B. Contemporary Machine Learning: A Guide for Practitioners in the Physical Sciences. arXiv 2017, arXiv:1712.08523.
[CrossRef]
39. The Dendral Project (Chapter 15)—The Quest for Artificial Intelligence. Available online: https://www.cambridge.org/core/
books/abs/quest-for-artificial-intelligence/dendral-project/7791DA5FAAF8D57E4B27E4EE387758E1 (accessed on 14 July 2022).
40. Rediscovering Some Problems of Artificial Intelligence in the Context of Organic Chemistry—Digital Collections—National
Library of Medicine. Available online: https://collections.nlm.nih.gov/catalog/nlm:nlmuid-101584906X921-doc (accessed on 25
August 2022).
41. Weiss, S.; Kulikowski, C.A.; Safir, A. Glaucoma Consultation by Computer. Comput. Biol. Med. 1978, 8, 25–40. [CrossRef]
42. Miller, R.A.; Pople, H.E.; Myers, J.D. Internist-I, an Experimental Computer-Based Diagnostic Consultant for General Internal
Medicine. N. Engl. J. Med. 1982, 307, 468–476. [CrossRef]
43. Sutton, R.T.; Pincock, D.; Baumgart, D.C.; Sadowski, D.C.; Fedorak, R.N.; Kroeker, K.I. An Overview of Clinical Decision Support
Systems: Benefits, Risks, and Strategies for Success. NPJ Digit. Med. 2020, 3, 1–10. [CrossRef] [PubMed]
44. Bone Tumor Diagnosis. Available online: http://uwmsk.org/bayes/bonetumor.html (accessed on 29 July 2022).
45. Lodwick, G.S.; Haun, C.L.; Smith, W.E.; Keller, R.F.; Robertson, E.D. Computer Diagnosis of Primary Bone Tumors. Radiology
1963, 80, 273–275. [CrossRef]
46. Belson, W.A. A Technique for Studying the Effects of a Television Broadcast. J. R. Stat. Soc. Ser. C (Appl. Stat.) 1956, 5, 195–202.
[CrossRef]
47. de Ville, B. Decision Trees. WIREs Comput. Stat. 2013, 5, 448–455. [CrossRef]
48. Belson, W.A. Matching and Prediction on the Principle of Biological Classification. J. R. Stat. Soc. Ser. C 1959, 8, 65–75. [CrossRef]
49. Ritschard, G. CHAID and Earlier Supervised Tree Methods; Routledge/Taylor & Francis Group: London, UK, 2010; p. 74;
ISBN 978-0-415-81706-6.
50. Morgan, J.N.; Sonquist, J.A. Problems in the Analysis of Survey Data, and a Proposal. J. Am. Stat. Assoc. 1963, 58, 415–434.
[CrossRef]
51. Gini, C. Variabilità e Mutabilità: Contributo allo Studio delle Distribuzioni e delle Relazioni Statistiche; Tipogr. di P. Cuppini: Bologna,
Italy, 1912. Available online: https://books.google.it/books?id=fqjaBPMxB9kC (accessed on 10 July 2022).
52. Podgorelec, V.; Kokol, P.; Stiglic, B.; Rozman, I. Decision Trees: An Overview and Their Use in Medicine. J. Med. Syst. 2002, 26,
445–463. [CrossRef]
53. Breiman, L.; Friedman, J.; Olshen, R.A.; Stone, C.J. Classification and Regression Trees. Available online: https://www.
taylorfrancis.com/books/mono/10.1201/9781315139470/classification-regression-trees-leo-breiman-jerome-friedman-
richard-olshen-charles-stone (accessed on 10 July 2022).
54. Shim, E.J.; Yoon, M.A.; Yoo, H.J.; Chee, C.G.; Lee, M.H.; Lee, S.H.; Chung, H.W.; Shin, M.J. An MRI-Based Decision Tree
to Distinguish Lipomas and Lipoma Variants from Well-Differentiated Liposarcoma of the Extremity and Superficial Trunk:
Classification and Regression Tree (CART) Analysis. Eur. J. Radiol. 2020, 127, 109012. [CrossRef]
Cancers 2024, 16, 3702 17 of 23

55. Convex Hull—An Overview. ScienceDirect Topics. Available online: https://www.sciencedirect.com/topics/mathematics/

convex-hull (accessed on 20 July 2022).
56. Vapnik, V.N.; Chervonenkis, A.Y. A Class of Algorithms for Pattern Recognition Learning. Avtomat. Telemekh. 1964, 25, 937–945.
Available online: http://www.mathnet.ru/php/archive.phtml?wshow=paper&jrnid=at&paperid=11678&option_lang=eng
(accessed on 22 July 2022).
57. Boser, B.E.; Guyon, I.M.; Vapnik, V.N. A Training Algorithm for Optimal Margin Classifiers. In Proceedings of the Proceedings of
the Fifth Annual Workshop on Computational Learning Theory, Pittsburgh, PA, USA, 27–29 July 1992; Association for Computing
Machinery: New York, NY, USA, 1992; pp. 144–152.
58. Guido, R.; Ferrisi, S.; Lofaro, D.; Conforti, D. An Overview on the Advancements of Support Vector Machine Models in Healthcare
Applications: A Review. Information 2024, 15, 235. [CrossRef]
59. Xing, X.; Zhao, X.; Wei, H.; Li, Y. Diagnostic Accuracy of Different Computer-Aided Diagnostic Systems for Prostate Cancer Based
on Magnetic Resonance Imaging. Medicine 2021, 100, e23817. [CrossRef] [PubMed]
60. Retico, A.; Gori, I.; Giuliano, A.; Muratori, F.; Calderoni, S. One-Class Support Vector Machines Identify the Language and
Default Mode Regions As Common Patterns of Structural Alterations in Young Children with Autism Spectrum Disorders. Front.
Neurosci. 2016, 10, 306. [CrossRef] [PubMed]
61. Retico, A.; Bosco, P.; Cerello, P.; Fiorina, E.; Chincarini, A.; Fantacci, M.E. Predictive Models Based on Support Vector Machines:
Whole-Brain versus Regional Analysis of Structural MRI in the Alzheimer’s Disease. J. Neuroimaging 2015, 25, 552–563. [CrossRef]
[PubMed]
62. Berrar, D. Bayes’ Theorem and Naive Bayes Classifier; Elsevier: Amsterdam, The Netherlands, 2019; pp. 403–412.
63. Kepka, J. The Current Approaches in Pattern Recognition. Kybernetika 1994, 30, 159–176.
64. Cover, T.; Hart, P. Nearest Neighbor Pattern Classification. IEEE Trans. Inf. Theory 1967, 13, 21–27. [CrossRef]
65. Fix, E.; Hodges, J.L. Discriminatory Analysis: Nonparametric Discrimination, Consistency Properties; USAF School of Aviation
Medicine: Randolph Field, TX, USA, 1951.
66. Hassani, C.; Varghese, B.A.; Nieva, J.; Duddalwar, V. Radiomics in Pulmonary Lesion Imaging. AJR Am. J. Roentgenol. 2019, 212,
497–504. [CrossRef]
67. Blum, A.L.; Langley, P. Selection of Relevant Features and Examples in Machine Learning. Artif. Intell. 1997, 97, 245–271.
[CrossRef]
68. Hall, M. Correlation-Based Feature Selection for Machine Learning. Ph.D. Thesis, The University of Waikato, Hamilton, New
Zealand, 1999.
69. Huang, S.H. Supervised Feature Selection: A Tutorial. Artif. Intell. Res. 2015, 4, 22. [CrossRef]
70. Esmeir, S.; Markovitch, S. Occam’s Razor Just Got Sharper. In Proceedings of the 20th International Joint Conference on Artificial
Intelligence, Hyderabad, India, 6–12 January 2007; pp. 768–773.
71. Probst, P.; Wright, M.N.; Boulesteix, A.-L. Hyperparameters and Tuning Strategies for Random Forest. WIREs Data Min. Knowl.
Discov. 2019, 9, e1301. [CrossRef]
72. Akaike, H. A New Look at the Statistical Model Identification. IEEE Trans. Autom. Control 1974, 19, 716–723. [CrossRef]
73. Bergstra, J.; Bengio, Y. Random Search for Hyper-Parameter Optimization. J. Mach. Learn. Res. 2012, 13, 281–305.
74. Fukushima, K. Neocognitron: A Self-Organizing Neural Network Model for a Mechanism of Pattern Recognition Unaffected by
Shift in Position. Biol. Cybern. 1980, 36, 193–202. [CrossRef] [PubMed]
75. Fukushima, K. Neocognitron: A Hierarchical Neural Network Capable of Visual Pattern Recognition. Neural Netw. 1988, 1,
119–130. [CrossRef]
76. Dechter, R. Learning While Searching in Constraint-Satisfaction-Problems. In Proceedings of the Fifth AAAI National Conference
on Artificial Intelligence, Philadelphia, PA, USA, 11–15 August 1986; pp. 178–183.
77. Fradkov, A.L. Early History of Machine Learning. IFAC-PapersOnLine 2020, 53, 1385–1390. [CrossRef]
78. Rumelhart, D.; Hinton, G.E.; Williams, R.J. Learning Representations by Back-Propagating Errors. Nature 1986, 323, 533–536.
[CrossRef]
79. Nielsen, M.A. Neural Networks and Deep Learning; Determination Press: San Francisco, CA, USA, 2015.
80. LeCun, Y.; Bengio, Y.; Hinton, G. Deep Learning. Nature 2015, 521, 436–444. [CrossRef]
81. Kononenko, I. Machine Learning for Medical Diagnosis: History, State of the Art and Perspective. Artif. Intell. Med. 2001, 23,
89–109. [CrossRef]
82. Glorot, X.; Bordes, A.; Bengio, Y. Deep Sparse Rectifier Neural Networks. In Proceedings of the Fourteenth International
Conference on Artificial Intelligence and Statistics, JMLR Workshop and Conference Proceedings, Fort Lauderdale, FL, USA,
11–13 April 2011; pp. 315–323.
83. Simard, P.Y.; Steinkraus, D.; Platt, J.C. Best Practices for Convolutional Neural Networks Applied to Visual Document Analysis.
In Proceedings of the Seventh International Conference on Document Analysis and Recognition, Edinburgh, UK, 3–6 August
2003; pp. 958–963.
84. Breiman, L. Bagging Predictors. Mach. Learn. 1996, 24, 123–140. [CrossRef]
85. Breiman, L. Random Forests. Mach. Learn. 2001, 45, 5–32. [CrossRef]
86. Efron, B. Bootstrap Methods: Another Look at the Jackknife. Ann. Stat. 1979, 7, 1–26. [CrossRef]
Cancers 2024, 16, 3702 18 of 23

87. Avanzo, M.; Pirrone, G.; Mileto, M.; Massarut, S.; Stancanello, J.; Baradaran-Ghahfarokhi, M.; Rink, A.; Barresi, L.; Vinante, L.;
Piccoli, E.; et al. Prediction of Skin Dose in Low-kV Intraoperative Radiotherapy Using Machine Learning Models Trained on
Results of in Vivo Dosimetry. Med. Phys. 2019, 46, 1447–1454. [CrossRef] [PubMed]
88. Ho, T.K. Random Decision Forests. In Proceedings of the Third International Conference on Document Analysis and Recognition,
Montreal, QC, Canada, 14–16 August 1995; p. 278.
89. Schapire, R.E. The Strength of Weak Learnability. Mach. Learn. 1990, 5, 197–227. [CrossRef]
90. Freund, Y.; Schapire, R.E. A Desicion-Theoretic Generalization of on-Line Learning and an Application to Boosting. In Proceedings
of the Computational Learning Theory; Vitányi, P., Ed.; Springer: Berlin/Heidelberg, Germany, 1995; pp. 23–37.
91. Giger, M.L.; Chan, H.-P.; Boone, J. Anniversary Paper: History and Status of CAD and Quantitative Image Analysis: The Role of
Medical Physics and AAPM. Med. Phys. 2008, 35, 5799–5820. [CrossRef] [PubMed]
92. Doi, K.; Giger, M.L.; Nishikawa, R.M.; Schmidt, R.A. Computer Aided Diagnosis of Breast Cancer on Mammograms. Breast
Cancer 1997, 4, 228–233. [CrossRef]
93. Le, E.P.V.; Wang, Y.; Huang, Y.; Hickman, S.; Gilbert, F.J. Artificial Intelligence in Breast Imaging. Clin. Radiol. 2019, 74, 357–366.
[CrossRef]
94. Ackerman, L.V.; Gose, E.E. Breast Lesion Classification by Computer and Xeroradiograph. Cancer 1972, 30, 1025–1035. [CrossRef]
95. Asada, N.; Doi, K.; MacMahon, H.; Montner, S.M.; Giger, M.L.; Abe, C.; Wu, Y. Potential Usefulness of an Artificial Neural
Network for Differential Diagnosis of Interstitial Lung Diseases: Pilot Study. Radiology 1990, 177, 857–860. [CrossRef]
96. U. S. Food and Drug Administration. Summary of Safety and Effectiveness Data: R2 Technologies (P970058). 1998. Available
online: https://www.accessdata.fda.gov/cdrh_docs/pdf/p970058.pdf (accessed on 28 October 2024).
97. Gilbert Fiona, J.; Astley Susan, M.; Gillan Maureen, G.C.; Agbaje Olorunsola, F.; Wallis Matthew, G.; James, J.; Boggis Caroline,
R.M.; Duffy Stephen, W. Single Reading with Computer-Aided Detection for Screening Mammography. N. Engl. J. Med. 2008, 359,
1675–1684. [CrossRef]
98. Haralick, R.M.; Shanmugam, K.; Dinstein, I. Textural Features for Image Classification. IEEE Trans. Syst. Man Cybern. 1973, 3,
610–621. [CrossRef]
99. Cavouras, D.; Prassopoulos, P. Computer Image Analysis of Brain CT Images for Discriminating Hypodense Cerebral Lesions in
Children. Med. Inform. 1994, 19, 13–20. [CrossRef]
100. Gillies, R.J.; Anderson, A.R.; Gatenby, R.A.; Morse, D.L. The Biology Underlying Molecular Imaging in Oncology: From Genome
to Anatome and Back Again. Clin. Radiol. 2010, 65, 517–521. [CrossRef] [PubMed]
101. Proceedings of the 2010 World Molecular Imaging Congress, Kyoto, Japan, 8–11 September 2010. Mol. Imaging Biol. 2010, 12,
500–1636. [CrossRef]
102. Falk, M.; Hausmann, M.; Lukasova, E.; Biswas, A.; Hildenbrand, G.; Davidkova, M.; Krasavin, E.; Kleibl, Z.; Falkova, I.; Jezkova,
L.; et al. Determining Omics Spatiotemporal Dimensions Using Exciting New Nanoscopy Techniques to Assess Complex Cell
Responses to DNA Damage: Part B--Structuromics. Crit. Rev. Eukaryot. Gene Expr. 2014, 24, 225–247. [CrossRef]
103. Avanzo, M.; Gagliardi, V.; Stancanello, J.; Blanck, O.; Pirrone, G.; El Naqa, I.; Revelant, A.; Sartor, G. Combining Computed
Tomography and Biologically Effective Dose in Radiomics and Deep Learning Improves Prediction of Tumor Response to Robotic
Lung Stereotactic Body Radiation Therapy. Med. Phys. 2021, 48, 6257–6269. [CrossRef] [PubMed]
104. Gillies, R.J.; Kinahan, P.E.; Hricak, H. Radiomics: Images Are More than Pictures, They Are Data. Radiology 2016, 278, 563–577.
[CrossRef]
105. Quartuccio, N.; Marrale, M.; Laudicella, R.; Alongi, P.; Siracusa, M.; Sturiale, L.; Arnone, G.; Cutaia, G.; Salvaggio, G.; Midiri,
M.; et al. The Role of PET Radiomic Features in Prostate Cancer: A Systematic Review. Clin. Transl. Imaging 2021, 9, 579–588.
[CrossRef]
106. Ubaldi, L.; Valenti, V.; Borgese, R.F.; Collura, G.; Fantacci, M.E.; Ferrera, G.; Iacoviello, G.; Abbate, B.F.; Laruina, F.; Tripoli, A.;
et al. Strategies to Develop Radiomics and Machine Learning Models for Lung Cancer Stage and Histology Prediction Using
Small Data Samples. Phys. Med. 2021, 90, 13–22. [CrossRef]
107. Pirrone, G.; Matrone, F.; Chiovati, P.; Manente, S.; Drigo, A.; Donofrio, A.; Cappelletto, C.; Borsatti, E.; Dassie, A.; Bortolus, R.;
et al. Predicting Local Failure after Partial Prostate Re-Irradiation Using a Dosiomic-Based Machine Learning Model. J. Pers. Med.
2022, 12, 1491. [CrossRef]
108. Avanzo, M.; Stancanello, J.; Pirrone, G.; Sartor, G. Radiomics and Deep Learning in Lung Cancer. Strahlenther. Onkol. 2020, 196,
879–887. [CrossRef]
109. Peira, E.; Sensi, F.; Rei, L.; Gianeri, R.; Tortora, D.; Fiz, F.; Piccardo, A.; Bottoni, G.; Morana, G.; Chincarini, A. Towards an
Automated Approach to the Semi-Quantification of [18 F]F-DOPA PET in Pediatric-Type Diffuse Gliomas. J. Clin. Med. 2023,
12, 2765. [CrossRef]
110. Ubaldi, L.; Saponaro, S.; Giuliano, A.; Talamonti, C.; Retico, A. Deriving Quantitative Information from Multiparametric MRI via
Radiomics: Evaluation of the Robustness and Predictive Value of Radiomic Features in the Discrimination of Low-Grade versus
High-Grade Gliomas with Machine Learning. Phys. Med. 2023, 107, 102538. [CrossRef] [PubMed]
111. Traverso, A.; Wee, L.; Dekker, A.; Gillies, R. Repeatability and Reproducibility of Radiomic Features: A Systematic Review. Int. J.
Radiat. Oncol. *Biol. *Phys. 2018, 102, 1143–1158. [CrossRef] [PubMed]
112. LeCun, Y.; Boser, B.; Denker, J.S.; Henderson, D.; Howard, R.E.; Hubbard, W.; Jackel, L.D. Backpropagation Applied to
Handwritten Zip Code Recognition. Neural Comput. 1989, 1, 541–551. [CrossRef]
Cancers 2024, 16, 3702 19 of 23

113. Srivastava, N.; Hinton, G.; Krizhevsky, A.; Sutskever, I.; Salakhutdinov, R. Dropout: A Simple Way to Prevent Neural Networks
from Overfitting. J. Mach. Learn. Res. 2014, 15, 1929–1958.
114. Krizhevsky, A.; Sutskever, I.; Hinton, G.E. ImageNet Classification with Deep Convolutional Neural Networks. In Proceedings of
the Advances in Neural Information Processing Systems; Curran Associates, Inc.: Red Hook, NY, USA, 2012; Volume 25.
115. Reinke, A.; Tizabi, M.D.; Eisenmann, M.; Maier-Hein, L. Common Pitfalls and Recommendations for Grand Challenges in Medical
Artificial Intelligence. Eur. Urol. Focus 2021, 7, 710–712. [CrossRef]
116. Castiglioni, I.; Rundo, L.; Codari, M.; Di Leo, G.; Salvatore, C.; Interlenghi, M.; Gallivanone, F.; Cozzi, A.; D’Amico, N.C.;
Sardanelli, F. AI Applications to Medical Images: From Machine Learning to Deep Learning. Phys. Med. 2021, 83, 9–24. [CrossRef]
117. Yu, K.-H.; Beam, A.L.; Kohane, I.S. Artificial Intelligence in Healthcare. Nat. Biomed. Eng. 2018, 2, 719–731. [CrossRef]
118. Chan, H.-P.; Hadjiiski, L.M.; Samala, R.K. Computer-Aided Diagnosis in the Era of Deep Learning. Med. Phys. 2020, 47, e218–e227.
[CrossRef]
119. Fujita, H. AI-Based Computer-Aided Diagnosis (AI-CAD): The Latest Review to Read First. Radiol. Phys. Technol. 2020, 13, 6–19.
[CrossRef]
120. Wu, Y.C.; Doi, K.; Giger, M.L. Detection of Lung Nodules in Digital Chest Radiographs Using Artificial Neural Networks: A Pilot
Study. J. Digit. Imaging 1995, 8, 88–94. [CrossRef]
121. Lo, S.C.; Lou, S.L.; Lin, J.S.; Freedman, M.T.; Chien, M.V.; Mun, S.K. Artificial Convolution Neural Network Techniques and
Applications for Lung Nodule Detection. IEEE Trans. Med. Imaging 1995, 14, 711–718. [CrossRef] [PubMed]
122. Lo, S.-C.B.; Lin, J.-S.; Freedman, M.T.; Mun, S.K. Computer-Assisted Diagnosis of Lung Nodule Detection Using Artificial
Convoultion Neural Network. In Proceedings of the Medical Imaging 1993: Image Processing, Newport Beach, CA, USA, 14–19
February 1993; Volume 1898, pp. 859–869.
123. Chan, H.P.; Lo, S.C.; Sahiner, B.; Lam, K.L.; Helvie, M.A. Computer-Aided Detection of Mammographic Microcalcifications:
Pattern Recognition with an Artificial Neural Network. Med. Phys. 1995, 22, 1555–1567. [CrossRef] [PubMed]
124. Zhou, S.K.; Greenspan, H.; Davatzikos, C.; Duncan, J.S.; Van Ginneken, B.; Madabhushi, A.; Prince, J.L.; Rueckert, D.; Summers,
R.M. A Review of Deep Learning in Medical Imaging: Imaging Traits, Technology Trends, Case Studies With Progress Highlights,
and Future Promises. Proc. IEEE 2021, 109, 820–838. [CrossRef] [PubMed]
125. Kaul, V.; Enslin, S.; Gross, S.A. History of Artificial Intelligence in Medicine. Gastrointest. Endosc. 2020, 92, 807–812. [CrossRef]
126. Lipkova, J.; Chen, R.J.; Chen, B.; Lu, M.Y.; Barbieri, M.; Shao, D.; Vaidya, A.J.; Chen, C.; Zhuang, L.; Williamson, D.F.; et al.
Artificial Intelligence for Multimodal Data Integration in Oncology. Cancer Cell 2022, 40, 1095–1110. [CrossRef]
127. Eixelberger, T.; Wolkenstein, G.; Hackner, R.; Bruns, V.; Mühldorfer, S.; Geissler, U.; Belle, S.; Wittenberg, T. YOLO Networks for
Polyp Detection: A Human-in-the-Loop Training Approach. Curr. Dir. Biomed. Eng. 2022, 8, 277–280. [CrossRef]
128. Ragab, M.G.; Abdulkadir, S.J.; Muneer, A.; Alqushaibi, A.; Sumiea, E.H.; Qureshi, R.; Al-Selwi, S.M.; Alhussian, H. A Comprehen-
sive Systematic Review of YOLO for Medical Object Detection (2018 to 2023). IEEE Access 2024, 12, 57815–57836. [CrossRef]
129. Gong, D.; Wu, L.; Zhang, J.; Mu, G.; Shen, L.; Liu, J.; Wang, Z.; Zhou, W.; An, P.; Huang, X.; et al. Detection of Colorectal
Adenomas with a Real-Time Computer-Aided System (ENDOANGEL): A Randomised Controlled Study. Lancet Gastroenterol.
Hepatol. 2020, 5, 352–361. [CrossRef]
130. Wang, Y.; Lombardo, E.; Avanzo, M.; Zschaek, S.; Weingärtner, J.; Holzgreve, A.; Albert, N.L.; Marschner, S.; Fanetti, G.; Franchin,
G.; et al. Deep Learning Based Time-to-Event Analysis with PET, CT and Joint PET/CT for Head and Neck Cancer Prognosis.
Comput. Methods Programs Biomed. 2022, 222, 106948. [CrossRef]
131. Perkins, D.; Salomon, G. Transfer of Learning. In The International Encyclopedia of Education, 2nd ed.; Husén, T., Postlethwaite,
T.N., Eds.; Pergamon: Oxford, UK, 1994; pp. 425–441.
132. Kim, H.E.; Cosa-Linan, A.; Santhanam, N.; Jannesari, M.; Maros, M.E.; Ganslandt, T. Transfer Learning for Medical Image
Classification: A Literature Review. BMC Med. Imaging 2022, 22, 69. [CrossRef]
133. Weiss, K.; Khoshgoftaar, T.M.; Wang, D. A Survey of Transfer Learning. J. Big Data 2016, 3, 9. [CrossRef]
134. Michael, E.; Ma, H.; Li, H.; Kulwa, F.; Li, J. Breast Cancer Segmentation Methods: Current Status and Future Potentials. Biomed.
Res. Int. 2021, 2021, 9962109. [CrossRef] [PubMed]
135. O’Donnell, M.; Gore, J.C.; Adams, W.J. Toward an Automated Analysis System for Nuclear Magnetic Resonance Imaging. II.
Initial Segmentation Algorithm. Med. Phys. 1986, 13, 293–297. [CrossRef] [PubMed]
136. Bezdek, J.C.; Hall, L.O.; Clarke, L.P. Review of MR Image Segmentation Techniques Using Pattern Recognition. Med. Phys. 1993,
20, 1033–1048. [CrossRef] [PubMed]
137. Comelli, A.; Stefano, A.; Russo, G.; Bignardi, S.; Sabini, M.G.; Petrucci, G.; Ippolito, M.; Yezzi, A. K-Nearest Neighbor Driving
Active Contours to Delineate Biological Tumor Volumes. Eng. Appl. Artif. Intell. 2019, 81, 133–144. [CrossRef]
138. Ronneberger, O.; Fischer, P.; Brox, T. U-Net: Convolutional Networks for Biomedical Image Segmentation; Navab, N., Hornegger, J.,
Wells, W.M., Frangi, A.F., Eds.; Springer International Publishing: Cham, Switzerland, 2015; pp. 234–241.
139. Oktay, O.; Schlemper, J.; Folgoc, L.L.; Lee, M.; Heinrich, M.; Misawa, K.; Mori, K.; McDonagh, S.; Hammerla, N.Y.; Kainz, B.; et al.
Attention U-Net: Learning Where to Look for the Pancreas. arXiv, 2018; arXiv:1804.03999. [CrossRef]
140. Lizzi, F.; Agosti, A.; Brero, F.; Cabini, R.F.; Fantacci, M.E.; Figini, S.; Lascialfari, A.; Laruina, F.; Oliva, P.; Piffer, S.; et al.
Quantification of Pulmonary Involvement in COVID-19 Pneumonia by Means of a Cascade of Two U-Nets: Training and
Assessment on Multiple Datasets Using Different Annotation Criteria. Int. J. CARS 2022, 17, 229–237. [CrossRef]
Cancers 2024, 16, 3702 20 of 23

141. Li, X.; Chen, H.; Qi, X.; Dou, Q.; Fu, C.-W.; Heng, P.A. H-DenseUNet: Hybrid Densely Connected UNet for Liver and Tumor
Segmentation from CT Volumes. IEEE Trans. Med. Imaging 2018, 37, 2663–2674. [CrossRef]
142. Khaled, R.; Vidal, J.; Vilanova, J.C.; Martí, R. A U-Net Ensemble for Breast Lesion Segmentation in DCE MRI. Comput. Biol. Med.
2022, 140, 105093. [CrossRef]
143. Moradi, S.; Oghli, M.G.; Alizadehasl, A.; Shiri, I.; Oveisi, N.; Oveisi, M.; Maleki, M.; Dhooge, J. MFP-Unet: A Novel Deep Learning
Based Approach for Left Ventricle Segmentation in Echocardiography. Phys. Medica 2019, 67, 58–69. [CrossRef]
144. Yi, X.; Walia, E.; Babyn, P. Generative Adversarial Network in Medical Imaging: A Review. Med. Image Anal. 2019, 58, 101552.
[CrossRef]
145. Singh, S.P.; Wang, L.; Gupta, S.; Goli, H.; Padmanabhan, P.; Gulyás, B. 3D Deep Learning on Medical Images: A Review. Sensors
2020, 20, 5097. [CrossRef] [PubMed]
146. Çiçek, Ö.; Abdulkadir, A.; Lienkamp, S.S.; Brox, T.; Ronneberger, O. 3D U-Net: Learning Dense Volumetric Segmentation from
Sparse Annotation. In Proceedings of the Medical Image Computing and Computer-Assisted Intervention—MICCAI 2016; Ourselin, S.,
Joskowicz, L., Sabuncu, M.R., Unal, G., Wells, W., Eds.; Springer International Publishing: Cham, Swizerland, 2016; pp. 424–432.
147. Goodfellow, I.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville, A.; Bengio, Y. Generative Adversarial
Nets. In Proceedings of the Advances in Neural Information Processing Systems; Curran Associates, Inc.: Red Hook, NY, USA, 2014;
Volume 27.
148. Shavlokhova, V.; Vollmer, A.; Zouboulis, C.C.; Vollmer, M.; Wollborn, J.; Lang, G.; Kübler, A.; Hartmann, S.; Stoll, C.; Roider, E.;
et al. Finetuning of GLIDE Stable Diffusion Model for AI-Based Text-Conditional Image Synthesis of Dermoscopic Images. Front.
Med. 2023, 10, 1231436. [CrossRef] [PubMed]
149. Toda, R.; Teramoto, A.; Tsujimoto, M.; Toyama, H.; Imaizumi, K.; Saito, K.; Fujita, H. Synthetic CT Image Generation of Shape-
Controlled Lung Cancer Using Semi-Conditional InfoGAN and Its Applicability for Type Classification. Int. J. Comput. Assist.
Radiol. Surg. 2021, 16, 241–251. [CrossRef] [PubMed]
150. Chlap, P.; Min, H.; Vandenberg, N.; Dowling, J.; Holloway, L.; Haworth, A. A Review of Medical Image Data Augmentation
Techniques for Deep Learning Applications. J. Med. Imaging Radiat. Oncol. 2021, 65, 545–563. [CrossRef]
151. Wolterink, J.M.; Mukhopadhyay, A.; Leiner, T.; Vogl, T.J.; Bucher, A.M.; Išgum, I. Generative Adversarial Networks: A Primer for
Radiologists. RadioGraphics 2021, 41, 840–857. [CrossRef]
152. Acosta, J.N.; Falcone, G.J.; Rajpurkar, P.; Topol, E.J. Multimodal Biomedical AI. Nat. Med. 2022, 28, 1773–1784. [CrossRef]
153. Gong, Y.; Shan, H.; Teng, Y.; Tu, N.; Li, M.; Liang, G.; Wang, G.; Wang, S. Parameter-Transferred Wasserstein Generative
Adversarial Network (PT-WGAN) for Low-Dose PET Image Denoising. IEEE Trans. Radiat. Plasma Med. Sci. 2021, 5, 213–223.
[CrossRef]
154. Lee, J. A Review of Deep Learning-Based Approaches for Attenuation Correction in Positron Emission Tomography. IEEE Trans.
Radiat. Plasma Med. Sci. 2020, 5, 160–184. [CrossRef]
155. Liu, M.; Zhu, A.H.; Maiti, P.; Thomopoulos, S.I.; Gadewar, S.; Chai, Y.; Kim, H.; Jahanshad, N.; for the Alzheimer’s Disease
Neuroimaging Initiative. Style Transfer Generative Adversarial Networks to Harmonize Multisite MRI to a Single Reference
Image to Avoid Overcorrection. Hum. Brain Mapp. 2023, 44, 4875–4892. [CrossRef]
156. Eo, T.; Jun, Y.; Kim, T.; Jang, J.; Lee, H.-J.; Hwang, D. KIKI-Net: Cross-Domain Convolutional Neural Networks for Reconstructing
Undersampled Magnetic Resonance Images. Magn. Reson. Med. 2018, 80, 2188–2201. [CrossRef]
157. Zhao, X.; Yang, T.; Li, B. A Review on Generative Based Methods for MRI Reconstruction. J. Phys. Conf. Ser. 2022, 2330, 012002.
[CrossRef]
158. Sloan, J.M.; Goatman, K.A.; Siebert, J.P. Learning Rigid Image Registration—Utilizing Convolutional Neural Networks for
Medical Image Registration. In Proceedings of the 11th International Joint Conference on Biomedical Engineering Systems and
Technologies, Funchal, Madeira, Portugal, 19–21 January 2018; pp. 89–99.
159. Fourcade, C.; Ferrer, L.; Moreau, N.; Santini, G.; Brennan, A.; Rousseau, C.; Lacombe, M.; Fleury, V.; Colombié, M.; Jézéquel, P.;
et al. Deformable Image Registration with Deep Network Priors: A Study on Longitudinal PET Images. Phys. Med. Biol. 2022,
67, 155011. [CrossRef] [PubMed]
160. Avanzo, M.; Barbiero, S.; Trovo, M.; Bissonnette, J.P.; Jena, R.; Stancanello, J.; Pirrone, G.; Matrone, F.; Minatel, E.; Cappelletto, C.;
et al. Voxel-by-Voxel Correlation between Radiologically Radiation Induced Lung Injury and Dose after Image-Guided, Intensity
Modulated Radiotherapy for Lung Tumors. Phys. Med. 2017, 42, 150–156. [CrossRef] [PubMed]
161. Xie, Y.; Takikawa, T.; Saito, S.; Litany, O.; Yan, S.; Khan, N.; Tombari, F.; Tompkin, J.; Sitzmann, V.; Sridhar, S. Neural Fields in
Visual Computing and Beyond. Comput. Graph. Forum 2022, 41, 641–676. [CrossRef]
162. Mao, S.; Sejdic, E. A Review of Recurrent Neural Network-Based Methods in Computational Physiology. IEEE Trans. Neural Netw.
Learn. Syst. 2023, 34, 6983–7003. [CrossRef]
163. Majumdar, A. and Gupta M. Recurrent Transfer Learning. Neural. Netw. 2019, 118, 271–279. [CrossRef]
164. Hochreiter, S.; Schmidhuber, J. Long Short-Term Memory. Neural Comput. 1997, 9, 1735–1780. [CrossRef]
165. Gers, F.A.; Schmidhuber, J.; Cummins, F. Learning to Forget: Continual Prediction with LSTM. Neural Comput. 2000, 12, 2451–2471.
[CrossRef]
166. Greff, K.; Srivastava, R.K.; Koutník, J.; Steunebrink, B.R.; Schmidhuber, J. LSTM: A Search Space Odyssey. IEEE Trans. Neural
Netw. Learn. Syst. 2017, 28, 2222–2232. [CrossRef]
Cancers 2024, 16, 3702 21 of 23

167. Steinkamp, J.; Cook, T.S. Basic Artificial Intelligence Techniques: Natural Language Processing of Radiology Reports. Radiol. Clin.
N. Am. 2021, 59, 919–931. [CrossRef]
168. Kreimeyer, K.; Foster, M.; Pandey, A.; Arya, N.; Halford, G.; Jones, S.F.; Forshee, R.; Walderhaug, M.; Botsis, T. Natural Language
Processing Systems for Capturing and Standardizing Unstructured Clinical Information: A Systematic Review. J. Biomed. Inform.
2017, 73, 14–29. [CrossRef] [PubMed]
169. Gultepe, E.; Green, J.P.; Nguyen, H.; Adams, J.; Albertson, T.; Tagkopoulos, I. From Vital Signs to Clinical Outcomes for Patients
with Sepsis: A Machine Learning Basis for a Clinical Decision Support System. J. Am. Med. Inform. Assoc. 2013, 21, 315–325.
[CrossRef] [PubMed]
170. Ravuri, M.; Kannan, A.; Tso, G.J.; Amatriain, X. Learning from the Experts: From Expert Systems to Machine-Learned Diagnosis
Models. In Proceedings of the 3rd Machine Learning for Healthcare Conference, Palo Alto, CA, USA, 17–18 August 2018;
pp. 227–243.
171. Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A.N.; Kaiser, L.; Polosukhin, I. Attention Is All You Need.
arXiv 2023, arXiv:1706.03762v7. [CrossRef]
172. Bajaj, S.; Gandhi, D.; Nayar, D. Potential Applications and Impact of ChatGPT in Radiology. Acad. Radiol. 2023, 31, 1256–1261.
[CrossRef]
173. Langlotz, C.P. The Future of AI and Informatics in Radiology: 10 Predictions. Radiology 2023, 309, e231114. [CrossRef]
174. Huang, J.; Neill, L.; Wittbrodt, M.; Melnick, D.; Klug, M.; Thompson, M.; Bailitz, J.; Loftus, T.; Malik, S.; Phull, A.; et al. Generative
Artificial Intelligence for Chest Radiograph Interpretation in the Emergency Department. JAMA Netw. Open 2023, 6, e2336100.
[CrossRef]
175. Ismail, A.; Ghorashi, N.S.; Javan, R. New Horizons: The Potential Role of OpenAI’s ChatGPT in Clinical Radiology. J. Am. Coll.
Radiol. 2023, 20, 696–698. [CrossRef]
176. Dosovitskiy, A.; Beyer, L.; Kolesnikov, A.; Weissenborn, D.; Zhai, X.; Unterthiner, T.; Dehghani, M.; Minderer, M.; Heigold, G.;
Gelly, S.; et al. An Image Is Worth 16x16 Words: Transformers for Image Recognition at Scale. arXiv 2021, arXiv:2010.11929.
[CrossRef]
177. Akbari, H.; Yuan, L.; Qian, R.; Chuang, W.-H.; Chang, S.-F.; Cui, Y.; Gong, B. VATT: Transformers for Multimodal Self-Supervised
Learning from Raw Video, Audio and Text. arXiv 2021, arXiv:2104.11178. [CrossRef]
178. Oren, O.; Gersh, B.J.; Bhatt, D.L. Artificial Intelligence in Medical Imaging: Switching from Radiographic Pathological Data to
Clinically Meaningful Endpoints. Lancet Digit. Health 2020, 2, e486–e488. [CrossRef]
179. Hafezi-Nejad, N.; Trivedi, P. Foundation AI Models and Data Extraction from Unlabeled Radiology Reports: Navigating
Uncharted Territory. Radiology 2023, 308, e232308. [CrossRef] [PubMed]
180. Moor, M.; Banerjee, O.; Abad, Z.S.H.; Krumholz, H.M.; Leskovec, J.; Topol, E.J.; Rajpurkar, P. Foundation Models for Generalist
Medical Artificial Intelligence. Nature 2023, 616, 259–265. [CrossRef] [PubMed]
181. Bluethgen, C.; Chambon, P.; Delbrouck, J.-B.; van der Sluijs, R.; Połacin, M.; Zambrano Chaves, J.M.; Abraham, T.M.; Purohit, S.;
Langlotz, C.P.; Chaudhari, A.S. A Vision-Language Foundation Model for the Generation of Realistic Chest X-Ray Images. Nat.
Biomed. Eng. 2024. [CrossRef] [PubMed]
182. Chen, R.J.; Ding, T.; Lu, M.Y.; Williamson, D.F.K.; Jaume, G.; Song, A.H.; Chen, B.; Zhang, A.; Shao, D.; Shaban, M.; et al. Towards
a General-Purpose Foundation Model for Computational Pathology. Nat. Med. 2024, 30, 850–862. [CrossRef]
183. Fink, M.A.; Bischoff, A.; Fink, C.A.; Moll, M.; Kroschke, J.; Dulz, L.; Heußel, C.P.; Kauczor, H.-U.; Weber, T.F. Potential of ChatGPT
and GPT-4 for Data Mining of Free-Text CT Reports on Lung Cancer. Radiology 2023, 308, e231362. [CrossRef]
184. Schäfer, R.; Nicke, T.; Höfener, H.; Lange, A.; Merhof, D.; Feuerhake, F.; Schulz, V.; Lotz, J.; Kiessling, F. Overcoming Data Scarcity
in Biomedical Imaging with a Foundational Multi-Task Model. Nat. Comput. Sci. 2024, 4, 495–509. [CrossRef]
185. Chen, X.; Wang, X.; Zhang, K.; Fung, K.-M.; Thai, T.C.; Moore, K.; Mannel, R.S.; Liu, H.; Zheng, B.; Qiu, Y. Recent Advances and
Clinical Applications of Deep Learning in Medical Image Analysis. Med. Image Anal. 2022, 79, 102444. [CrossRef]
186. Silver, D.; Schrittwieser, J.; Simonyan, K.; Antonoglou, I.; Huang, A.; Guez, A.; Hubert, T.; Baker, L.; Lai, M.; Bolton, A.; et al.
Mastering the Game of Go without Human Knowledge. Nature 2017, 550, 354–359. [CrossRef]
187. Esteva, A.; Robicquet, A.; Ramsundar, B.; Kuleshov, V.; DePristo, M.; Chou, K.; Cui, C.; Corrado, G.; Thrun, S.; Dean, J. A Guide to
Deep Learning in Healthcare. Nat. Med. 2019, 25, 24–29. [CrossRef]
188. Antropova, N.; Huynh, B.Q.; Giger, M.L. A Deep Feature Fusion Methodology for Breast Cancer Diagnosis Demonstrated on
Three Imaging Modality Datasets. Med. Phys. 2017, 44, 5162–5171. [CrossRef]
189. Terpstra, M.L.; Maspero, M.; D’Agata, F.; Stemkens, B.; Intven, M.P.W.; Lagendijk, J.J.W.; Van den Berg, C.A.T.; Tijssen, R.H.N. Deep
Learning-Based Image Reconstruction and Motion Estimation from Undersampled Radial k-Space for Real-Time MRI-Guided
Radiotherapy. Phys. Med. Biol. 2020, 65, 155015. [CrossRef] [PubMed]
190. Furlong, J.W.; Dupuy, M.E.; Heinsimer, J.A. Neural Network Analysis of Serial Cardiac Enzyme Data A Clinical Application of
Artificial Machine Intelligence. Am. J. Clin. Pathol. 1991, 96, 134–141. [CrossRef]
191. Baxt, W.G. Use of an Artificial Neural Network for the Diagnosis of Myocardial Infarction. Ann. Intern. Med. 1991, 115, 843–848.
[CrossRef] [PubMed]
192. Gross, G.W.; Boone, J.M.; Greco-Hunt, V.; Greenberg, B. Neural Networks in Radiologic Diagnosis. II. Interpretation of Neonatal
Chest Radiographs. Investig. Radiol. 1990, 25, 1017–1023. [CrossRef] [PubMed]
Cancers 2024, 16, 3702 22 of 23

193. Romagnoni, A.; Jégou, S.; Van Steen, K.; Wainrib, G.; Hugot, J.-P. Comparative Performances of Machine Learning Methods for
Classifying Crohn Disease Patients Using Genome-Wide Genotyping Data. Sci. Rep. 2019, 9, 10351. [CrossRef] [PubMed]
194. Vântu, A.; Vasilescu, A.; Băicoianu, A. Medical Emergency Department Triage Data Processing Using a Machine-Learning
Solution. Heliyon 2023, 9, e18402. [CrossRef]
195. Momenzadeh, M.; Vard, A.; Talebi, A.; Mehri Dehnavi, A.; Rabbani, H. Computer-Aided Diagnosis Software for Vulvovaginal
Candidiasis Detection from Pap Smear Images. Microsc. Res. Tech. 2018, 81, 13–21. [CrossRef]
196. Girdhar, R.; El-Nouby, A.; Liu, Z.; Singh, M.; Alwala, K.V.; Joulin, A.; Misra, I. ImageBind One Embedding Space to Bind Them
All. In Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Vancouver, BC,
Canada, 18–22 June 2023; pp. 15180–15190.
197. Choi, K.-H.; Ha, J.-E. Semantic Segmentation with Perceiver IO. In Proceedings of the 2022 22nd International Conference on
Control, Automation and Systems (ICCAS), Jeju, Republic of Korea, 27 November–1 December 2022; pp. 1607–1610.
198. Mancosu, P.; Lambri, N.; Castiglioni, I.; Dei, D.; Iori, M.; Loiacono, D.; Russo, S.; Talamonti, C.; Villaggi, E.; Scorsetti, M.
Applications of artificial intelligence in stereotactic body radiation therapy. Phys. Med. Biol. 2022, 67. [CrossRef]
199. Robust machine learning challenge: An AIFM multicentric competition to spread knowledge, identify common pitfalls and
recommend best practice. Phys. Med. 2024, 127, 104834. [CrossRef]
200. Egger, J.; Gsaxner, C.; Pepe, A.; Pomykala, K.L.; Jonske, F.; Kurz, M.; Li, J.; Kleesiek, J. Medical Deep Learning—A Systematic
Meta-Review. Comput. Methods Programs Biomed. 2022, 221, 106874. [CrossRef]
201. Stollmayer, R.; Budai, B.K.; Rónaszéki, A.; Zsombor, Z.; Kalina, I.; Hartmann, E.; Tóth, G.; Szoldán, P.; Bérczi, V.; Maurovich-
Horvat, P.; et al. Focal Liver Lesion MRI Feature Identification Using Efficientnet and MONAI: A Feasibility Study. Cells 2022, 11,
1558. [CrossRef] [PubMed]
202. Gillot, M.; Baquero, B.; Le, C.; Deleat-Besson, R.; Bianchi, J.; Ruellas, A.; Gurgel, M.; Yatabe, M.; Turkestani, N.A.; Najarian, K.;
et al. Automatic Multi-Anatomical Skull Structure Segmentation of Cone-Beam Computed Tomography Scans Using 3D UNETR.
PLoS ONE 2022, 17, e0275033. [CrossRef] [PubMed]
203. Termine, A.; Fabrizio, C.; Caltagirone, C.; Petrosini, L.; on behalf of the Frontotemporal Lobar Degeneration Neuroimaging
Initiative. A Reproducible Deep-Learning-Based Computer-Aided Diagnosis Tool for Frontotemporal Dementia Using MONAI
and Clinica Frameworks. Life 2022, 12, 947. [CrossRef] [PubMed]
204. Vallieres, M.; Zwanenburg, A.; Badic, B.; Cheze Le Rest, C.; Visvikis, D.; Hatt, M. Responsible Radiomics Research for Faster
Clinical Translation. J. Nucl. Med. 2018, 59, 189–193. [CrossRef]
205. Zhang, S.; Liu, R.; Wang, Y.; Zhang, Y.; Li, M.; Wang, Y.; Wang, S.; Ma, N.; Ren, J. Ultrasound-Base Radiomics for Discerning
Lymph Node Metastasis in Thyroid Cancer: A Systematic Review and Meta-Analysis. Acad. Radiol. 2024, 31, 3118–3130.
[CrossRef]
206. Kocak, B.B.B.; Bakas, S.; Cuocolo, R.; Fedorov, A.; Maier-Hein, L.; Mercaldo, N.; Müller, H.; Orlhac, F.; Pinto Dos Santos, D.;
Stanzione, A.; et al. CheckList for EvaluAtion of Radiomics Research (CLEAR): A Step-by-Step Reporting Guideline for Authors
and Reviewers Endorsed by ESR and EuSoMII. Insights Imaging 2023, 14, 75. [CrossRef]
207. Acharya, U.R.; Hagiwara, Y.; Sudarshan, V.K.; Chan, W.Y.; Ng, K.H. Towards Precision Medicine: From Quantitative Imaging to
Radiomics. J. Zhejiang Univ. Sci. B 2018, 19, 6–24. [CrossRef]
208. Park, C.J.; Park, Y.W.; Ahn, S.S.; Kim, D.; Kim, E.H.; Kang, S.-G.; Chang, J.H.; Kim, S.H.; Lee, S.-K. Quality of Radiomics Research
on Brain Metastasis: A Roadmap to Promote Clinical Translation. Korean J. Radiol. 2022, 23, 77–88. [CrossRef]
209. Avery, E.W.; Behland, J.; Mak, A.; Haider, S.P.; Zeevi, T.; Sanelli, P.C.; Filippi, C.G.; Malhotra, A.; Matouk, C.C.; Griessenauer, C.J.;
et al. Dataset on Acute Stroke Risk Stratification from CT Angiographic Radiomics. Data Brief. 2022, 44, 108542. [CrossRef]
210. Prior, F.W.; Clark, K.; Commean, P.; Freymann, J.; Jaffe, C.; Kirby, J.; Moore, S.; Smith, K.; Tarbox, L.; Vendt, B.; et al. TCIA: An
Information Resource to Enable Open Science. Conf. Proc. IEEE Eng. Med. Biol. Soc. 2013, 2013, 1282–1285. [CrossRef]
211. van Leeuwen, K.G.; Schalekamp, S.; Rutten, M.J.C.M.; van Ginneken, B.; de Rooij, M. Artificial Intelligence in Radiology: 100
Commercially Available Products and Their Scientific Evidence. Eur. Radiol. 2021, 31, 3797–3804. [CrossRef] [PubMed]
212. Park, S.H.; Han, K.; Jang, H.Y.; Park, J.E.; Lee, J.-G.; Kim, D.W.; Choi, J. Methods for Clinical Evaluation of Artificial Intelligence
Algorithms for Medical Diagnosis. Radiology 2023, 306, 20–31. [CrossRef] [PubMed]
213. Wu, E.; Wu, K.; Daneshjou, R.; Ouyang, D.; Ho, D.E.; Zou, J. How Medical AI Devices Are Evaluated: Limitations and
Recommendations from an Analysis of FDA Approvals. Nat. Med. 2021, 27, 582–584. [CrossRef] [PubMed]
214. Han, R.; Acosta, J.N.; Shakeri, Z.; Ioannidis, J.P.A.; Topol, E.J.; Rajpurkar, P. Randomised Controlled Trials Evaluating Artificial
Intelligence in Clinical Practice: A Scoping Review. Lancet Digit. Health 2024, 6, e367–e373. [CrossRef] [PubMed]
215. Fazal, M.I.; Patel, M.E.; Tye, J.; Gupta, Y. The Past, Present and Future Role of Artificial Intelligence in Imaging. Eur. J. Radiol.
2018, 105, 246–250. [CrossRef]
216. Neri, E.; Aghakhanyan, G.; Zerunian, M.; Gandolfo, N.; Grassi, R.; Miele, V.; Giovagnoni, A.; Laghi, A.; SIRM expert group on
Artificial Intelligence. Explainable AI in Radiology: A White Paper of the Italian Society of Medical and Interventional Radiology.
Radiol. Med. 2023, 128, 755–764. [CrossRef]
217. Avanzo, M.; Pirrone, G.; Vinante, L.; Caroli, A.; Stancanello, J.; Drigo, A.; Massarut, S.; Mileto, M.; Urbani, M.; Trovo, M.;
et al. Electron Density and Biologically Effective Dose (BED) Radiomics-Based Machine Learning Models to Predict Late
Radiation-Induced Subcutaneous Fibrosis. Front. Oncol. 2020, 10, 490. [CrossRef]
Cancers 2024, 16, 3702 23 of 23

218. Murdoch, W.J.; Singh, C.; Kumbier, K.; Abbasi-Asl, R.; Yu, B. Definitions, Methods, and Applications in Interpretable Machine
Learning. Proc. Natl. Acad. Sci. USA 2019, 116, 22071–22080. [CrossRef]
219. Neves, J.; Hsieh, C.; Nobre, I.B.; Sousa, S.C.; Ouyang, C.; Maciel, A.; Duchowski, A.; Jorge, J.; Moreira, C. Shedding Light on Ai in
Radiology: A Systematic Review and Taxonomy of Eye Gaze-Driven Interpretability in Deep Learning. Eur. J. Radiol. 2024, 172,
111341. [CrossRef]
220. Champendal, M.; Müller, H.; Prior, J.O.; dos Reis, C.S. A Scoping Review of Interpretability and Explainability Concerning
Artificial Intelligence Methods in Medical Imaging. Eur. J. Radiol. 2023, 169, 111159. [CrossRef]
221. Ricci Lara, M.A.; Echeveste, R.; Ferrante, E. Addressing Fairness in Artificial Intelligence for Medical Imaging. Nat. Commun.
2022, 13, 4581. [CrossRef] [PubMed]
222. Burlina, P.; Joshi, N.; Paul, W.; Pacheco, K.D.; Bressler, N.M. Addressing Artificial Intelligence Bias in Retinal Diagnostics. Transl.
Vis. Sci. Technol. 2021, 10, 13. [CrossRef] [PubMed]
223. Mahmood, U.; Shukla-Dave, A.; Chan, H.P.; Drukker, K.; Samala, R.K.; Chen, Q.; Vergara, D.; Greenspan, H.; Petrick, N.; Sahiner,
B.; et al. Artificial Intelligence in Medicine: Mitigating Risks and Maximizing Benefits via Quality Assurance, Quality Control,
and Acceptance Testing. BJR|Artif. Intell. 2024, 1, ubae003. [CrossRef] [PubMed]
224. Kelly, B.S.; Quinn, C.; Belton, N.; Lawlor, A.; Killeen, R.P.; Burrell, J. Cybersecurity Considerations for Radiology Departments
Involved with Artificial Intelligence. Eur. Radiol. 2023, 33, 8833–8841. [CrossRef] [PubMed]
225. Mahadevaiah, G.; Rv, P.; Bermejo, I.; Jaffray, D.; Dekker, A.; Wee, L. Artificial Intelligence-Based Clinical Decision Support
in Modern Medical Physics: Selection, Acceptance, Commissioning, and Quality Assurance. Med. Phys. 2020, 47, e228–e235.
[CrossRef]
226. COCIR. COCIR Analysis on AI in Medical Device Legislation—May 2021. Available online: https://www.cocir.org/latest-news/
publications/article/cocir-analysis-on-ai-in-medical-device-legislation-may-2021 (accessed on 9 April 2024).
227. Ebers, M.; Hoch, V.R.S.; Rosenkranz, F.; Ruschemeier, H.; Steinrötter, B. The European Commission’s Proposal for an Artificial
Intelligence Act—A Critical Assessment by Members of the Robotics and AI Law Society (RAILS). J 2021, 4, 589–603. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.