NeuroImage: Clinical 36 (2022) 103205

Contents lists available at ScienceDirect

NeuroImage: Clinical
journal homepage: www.elsevier.com/locate/ynicl

Cortical lesions, central vein sign, and paramagnetic rim lesions in multiple
sclerosis: Emerging machine learning techniques and future avenues
Francesco La Rosa a, b, c, *, Maxence Wynen b, d, e, f, Omar Al-Louzi g, h, Erin S Beck c, g,
Till Huelnhagen a, f, i, Pietro Maggi e, j, k, Jean-Philippe Thiran a, b, f, Tobias Kober a, f, i, Russell
T Shinohara l, m, n, Pascal Sati g, h, Daniel S Reich g, Cristina Granziera o, p, Martina Absinta q, r,
Meritxell Bach Cuadra b, f
a
Signal Processing Laboratory (LTS5), Ecole Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland
b
CIBM Center for Biomedical Imaging, Switzerland
c
Department of Neurology, Icahn School of Medicine at Mount Sinai, New York, NY, USA
d
ICTeam, UCLouvain, Louvain-la-Neuve, Belgium
e
Louvain Inflammation Imaging Lab (NIL), Institute of Neuroscience (IoNS), UCLouvain, Brussels, Belgium
f
Radiology Department, Lausanne University and University Hospital, Switzerland
g
Translational Neuroradiology Section, National Institute of Neurological Disorders and Stroke, National Institutes of Health, Bethesda, MD, USA
h
Department of Neurology, Cedars-Sinai Medical Center, Los Angeles, CA, USA
i
Advanced Clinical Imaging Technology, Siemens Healthcare AG, Lausanne, Switzerland
j
Department of Neurology, Cliniques universitaires Saint-Luc, Université catholique de Louvain, Brussels, Belgium
k
Department of Neurology, CHUV, Lausanne, Switzerland
l
Center for Biomedical Image Computing and Analysis (CBICA), Department of Radiology, University of Pennsylvania, Philadelphia, PA, USA
m
Penn Statistics in Imaging and Visualization Endeavor (PennSIVE), Center for Clinical Epidemiology and Biostatistics, University of Pennsylvania, Philadelphia, PA,
USA
n
Department of Biostatistics, Epidemiology, and Informatics, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA
o
Translational Imaging in Neurology (ThINk) Basel, Department of Biomedical Engineering, Faculty of Medicine, University Hospital Basel and University of Basel,
Switzerland
p
Neurologic Clinic and Policlinic, MS Center and Research Center for Clinical Neuroimmunology and Neuroscience Basel (RC2NB), University Hospital Basel and
University of Basel, Basel, Switzerland
q
IRCCS San Raffaele Hospital and Vita-Salute San Raffaele University, Milan, Italy
r
Department of Neurology, Johns Hopkins University School of Medicine, Baltimore, MD, USA

A B S T R A C T

The current diagnostic criteria for multiple sclerosis (MS) lack specificity, and this may lead to misdiagnosis, which remains an issue in present-day clinical practice.
In addition, conventional biomarkers only moderately correlate with MS disease progression. Recently, some MS lesional imaging biomarkers such as cortical lesions
(CL), the central vein sign (CVS), and paramagnetic rim lesions (PRL), visible in specialized magnetic resonance imaging (MRI) sequences, have shown higher
specificity in differential diagnosis. Moreover, studies have shown that CL and PRL are potential prognostic biomarkers, the former correlating with cognitive
impairments and the latter with early disability progression. As machine learning-based methods have achieved extraordinary performance in the assessment of
conventional imaging biomarkers, such as white matter lesion segmentation, several automated or semi-automated methods have been proposed as well for CL, PRL,
and CVS. In the present review, we first introduce these MS biomarkers and their imaging methods. Subsequently, we describe the corresponding machine learning-
based methods that were proposed to tackle these clinical questions, putting them into context with respect to the challenges they are facing, including non-
standardized MRI protocols, limited datasets, and moderate inter-rater variability. We conclude by presenting the current limitations that prevent their broader
deployment and suggesting future research directions.

Abbreviations: MS, multiple sclerosis; MRI, magnetic resonance imaging; DL, deep learning; ML, machine learning; CL, cortical lesions; PRL, paramagnetic rim
lesions; CVS, central vein sign; WML, white matter lesions; FLAIR, fluid-attenuated inversion recovery; MPRAGE, magnetization prepared rapid gradient-echo; GM,
gray matter; WM, white matter; PSIR, phase-sensitive inversion recovery; DIR, double inversion recovery; MP2RAGE, magnetization-prepared 2 rapid gradient
echoes; SELs, Slowly evolving/expanding lesions; CNN, convolutional neural network; XAI, explainable AI; PV, partial volume.
* Corresponding author at: Department of Neurology, Icahn School of Medicine at Mount Sinai, New York, NY, USA.
E-mail address: francesco.larosa@mssm.edu (F. La Rosa).

https://doi.org/10.1016/j.nicl.2022.103205
Received 13 December 2021; Received in revised form 9 September 2022; Accepted 16 September 2022
Available online 24 September 2022
2213-1582/© 2022 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
F. La Rosa et al. NeuroImage: Clinical 36 (2022) 103205

1. Introduction segmentation and derived scalar measurements from diffusion-based

measurements (FA, MD, NODDI parameters, etc) averaged at the
Multiple sclerosis (MS) is a chronic inflammatory disease and a lesion level (Lu et al., 2021; Oladosu et al., 2021; Ye et al., 2020; Mar­
common cause of neurological disability in young adults (Reich et al., tínez-Heras et al., 2020). Further studies, however, are still needed to
2018). Its hallmark is demyelinated white matter lesions (WML) forming verify the possible use of these quantitative features for patient
in the central nervous system (Reich et al., 2018). These lesions are stratification.
assessed in-vivo with magnetic resonance imaging (MRI), which is the Recently, advances in MR technology, such as the development of
imaging technique of choice to diagnose MS and monitor the disease specialized sequences, acceleration of protocols, and the proliferation of
over time (Hemond and Bakshi, 2018). The current MRI diagnostic ultra-high field MRI, have allowed the imaging of pathologically specific
criteria (McDonald criteria) are based on the dissemination in space and MS lesional biomarkers (Cortese et al., 2019; Ineichen et al., 2021).
time of such lesions (Thompson et al., 2018). Moreover, the quantifi­ These include cortical lesions (CL), the central vein sign (CVS), and
cation of the total lesion volume is important to determine ongoing paramagnetic rim lesions (PRL). Studies have shown that CL and PRL are
disease activity and monitor treatment effect over time (Giorgio et al., potential prognostic biomarkers: CL are associated with cognitive im­
2014). Recommended MRI techniques include T2 and T1-weighted pairments, while patients with PRL experience an earlier progression in
inversion recovery sequences, such as fluid-attenuated inversion re­ disability (Calabrese et al., 2010; Absinta et al., 2019). Furthermore, the
covery (FLAIR), and magnetization prepared rapid gradient-echo CVS and PRL have proven to be effective for differentiating MS from
(MPRAGE) (Wattjes et al., 2021). At common clinical magnetic fields mimicking diseases (Ontaneda et al., 2021; Sati et al., 2016; Clarke et al.,
(1.5 T and 3 T), the use of gadolinium-based contrast agents is useful to 2020; Maggi et al., 2020). All three biomarkers, however, require
evaluate patients suspected of MS and monitor disease activity causing dedicated MRI sequences at high (3 T) or ultra-high (7 T) magnetic
breakdown of the blood–brain barrier (Filippi et al., 2019). fields, and experienced raters for their manual assessment, which can be
As the manual detection of WML is time-consuming and prone to very time-consuming. As done in the past for WML, various automated
inter-rater variability (Hagens et al., 2019), a myriad of automated or or semi-automated methods, mostly based on machine learning (ML),
semi-automated approaches have been developed to facilitate this task have been developed to facilitate the three aforementioned biomarkers’
(Lladó et al., 2012). These methods were initially based primarily on assessment (see Table 1). Compared to their WML counterparts, how­
MRI intensity features and probabilistic atlases (Lladó et al., 2012), ever, they face additional challenges, including non-standardized im­
whereas, more recently, the vast majority use deep learning (DL) ap­ aging protocols, moderate inter-rater variability when determining
proaches (Zeng et al., 2020), without prior feature extraction. Sub­ ground truth annotations, and smaller datasets. Automated assessment
stantial effort is now being made towards reproducibility of the results could improve standardization and facilitate large-scale assessment in
and open science (Vrenken et al., 2021). Several grand challenges have clinical routine of the aforementioned biomarkers, with clear benefits in
been organized (Carass et al., 2017; Commowick et al., 2018; Commo­ terms of MS diagnosis and prognosis.
wick et al., 2021), in which DL-based methods have achieved the best In this review, we first briefly describe these advanced imaging
performance, approaching or sometimes even outperforming human biomarkers and their imaging requirements and then focus on image
readers (Carass et al., 2017; Commowick et al., 2021). WML segmen­ processing techniques tailored for their automated segmentation and
tation methods have been reviewed recently (Zeng et al., 2020; Kaur classification. We conclude with a discussion on current limitations and
et al., 2021); the present review thus focuses on machine learning future lines of research to boost the development of ML approaches in
techniques tailored for lesional biomarkers specific to MS that require this area and encourage their adoption in MS research and clinical
advanced MRI techniques and have the potential to improve MS diag­ settings.
nosis and prognosis.
One major drawback of the current MS diagnostic criteria is their 2. Cortical lesions, paramagnetic lesions, and central vein sign
lack of specificity, as they were proposed to identify patients with a high
likelihood of MS rather than distinguish MS from other conditions In this section, we present a brief description of CL, CVS, and PRL,
(Thompson et al., 2018). The lack of specificity of these criteria may lead and their respective imaging protocols. In addition to the CVS and PRL,
to misdiagnosis, which remains a persistent problem of MS (Solomon which have emerged as promising MS biomarkers in recent years, we
et al., 2019). Multi-center studies have shown a misdiagnosis rate of also included CL which, although included in the MS diagnostic criteria,
18% (Kaisey et al., 2019), often associated with atypical clinical or are not yet commonly analyzed in clinical practice. For the sake of
imaging findings. Improving the diagnostic specificity would prevent completeness, a short description of slowly expanding lesions (SELs) is
harmful consequences for patients (Solomon et al., 2016) and allow also provided, although these have not been assessed with ML-based
clinicians to prescribe the appropriate treatment earlier. In addition, approaches yet.
although clinical relapses are often associated with the appearance of Cortical lesions (CL) - Cortical lesions are a type of MS lesions that
new WML, the overall WML burden, which is the most common MRI involve, at least partially, the cortex and have been classified into three
biomarker examined in clinical routine, is only moderately correlated main categories (Calabrese et al., 2010) (see Fig. 1): leukocortical le­
with disability and poorly predicts transition to progressive disease sions are located at the interface between WM and gray matter (GM)
(Barkhof, 2002). For all these reasons, there is a need for additional (type I), intracortical lesions are purely in the cortex and do not reach
biomarkers that are highly specific to MS or correlate with disease the pial surface (type II), and subpial lesions touch the subpial surface of
progression. the brain (type III) and sometimes extend all the way to the white matter
Quantitative MRI, such as relaxometry, myelin imaging, or diffusion (type IV). Cortical demyelination in MS has long been recognized in
MR, provides information related to the microstructural composition pathology studies, but only in the last two decades have dedicated se­
and organization of tissues. In MS, quantitative MRI techniques com­ quences on high- and ultra-high field scanners provided in-vivo evidence
plement conventional MRI techniques by providing insights into disease of cortical damage (Calabrese et al., 2010). Cortical lesions are clinically
mechanisms (Granziera et al., 2021). For instance, diffusion tensor im­ interestingfor several reasons. First, they have been observed in the
aging and microstructure models of diffusion can help better understand early stages of the disease and in all of the major MS phenotypes (Kidd
the MS lesion heterogeneity (myelin and axonal damage). Voxel-wise et al., 1999). Second, they are associated with disability (Harrison et al.,
analysis methods allow exploring group-wise differences without the 2015; Nielsen et al., 2013; Calabrese et al., 2012) and in some studies,
need for prior lesion segmentation (Thaler et al., 2021;16(2): their number was associated with cognitive disability more strongly
e0245844.). On the contrary, classification methods in this context have than the number of WML (Harrison et al., 2015; Favaretto et al., 2016).
been used to cluster different lesion types based on prior lesion Third, longitudinal studies have linked them with disease progression

2
F. La Rosa et al. NeuroImage: Clinical 36 (2022) 103205

(Treaba et al., 2019; Mainero et al., 2015; Scalfari et al., 2018; Calabrese be reliably observed across different T2* sequences at 3 T, although the
et al., 2013). Fourth, subpial cortical demyelination is highly specific to sensitivity depends on the sequence considered (Samaraweera et al.,
MS (Junker et al., 2020); CL have been observed in patients with 2017). To obtain the best detection sensitivity for the CVS, optimized
radiologically isolated syndrome (Giorgio et al., 2011), but not in pa­ MRI acquisitions have been proposed (T2*-weighted acquired with 3D-
tients with neuromyelitis optica (Sinnecker et al., 2012). Since 2017, CL segmented echo-planar-imaging or T2*w 3D-EPI (Sati et al., 2014),
have been included in the MS diagnostic criteria (Thompson et al., combined T2-FLAIR and T2*, also called FLAIR* (Sati et al., 2012), and
2018), but their visualization from routine MRI sequences remains susceptibility-based sequence, called SWAN-Venule (Gaitán et al.,
difficult. For instance, a postmortem study showed that 3D FLAIR at 3 T 2020). These sequences were shown to provide superior CVS detection
could detect about 41% of leukocortical lesions and only 5% of intra­ compared to clinical acquisitions at 1.5 T and 3 T (Castellaro et al., 2020;
cortical and subpial lesions (Geurts et al., 2005). This supports the need Suh et al., 2019). Single-center and multi-center retrospective studies
for specialized MRI techniques (see Fig. 2) such as the phase-sensitive imaging patients with clinically established diagnoses have demon­
inversion recovery (PSIR), double inversion recovery (DIR), and strated a significantly higher proportion of CVS-positive white matter
magnetization-prepared 2 rapid gradient echoes (MP2RAGE) (Filippi lesions (%CVS + ) in MS (mean pooled incidence: 79%, 95% CI:
et al., 2019; Müller et al., 2022). However, these sequences are still 68–87%) (Suh et al., 2019) as compared to other neurological disorders
relatively insensitive to CL at 1.5 T and 3 T (Müller et al., 2022; Kilsdonk mimicking MS (mean pooled incidence: 38%, 95% CI: 18–63%) (Suh
et al., 2016; Beck et al., 2020). Ultra-high field MRI, with its higher et al., 2019) such as cerebral small vessel disease (Campion et al., 2017),
signal-to-noise ratio and increased susceptibility effects, has proven to neuromyelitis optica spectrum disorder (NMOSD) (Cortese et al., 2018),
be a powerful tool for increasing the sensitivity to CL, especially for inflammatory vasculopathies (Maggi et al., 2018), and migraine (Solo­
intracortical and subpial lesions (Madsen et al., 2021; Beck et al., 2022; mon et al., 2016). To distinguish MS from other neurological conditions,
Maranzano et al., 2019). Even with the most sensitive methods, how­ different CVS-based criteria have been proposed to date, some based on
ever, CL are small and often subtle, making manual segmentation the percentage of perivenular lesions (from 35% to 60%) and others
extremely time consuming and subject to relatively low inter-rater simply on the CVS lesion count (3-lesion or 6-lesion rule) (Maggi et al.,
reliability (Harrison et al., 2015; Faizy et al., 2017). 2018; Tallantyre et al., 2011; Mistry et al., 2016; Solomon et al., 2018).
Central vein sign (CVS) - Recently, studies have suggested that an From a diagnostic perspective, retrospective studies have shown excel­
MRI-detectable central vein inside MS lesions might be evidence of lent diagnostic discrimination by applying the ‘40% rule’ (Tallantyre
pathological processes specific to MS (see Fig. 3) (Maggi et al., 2018; et al., 2011) with sensitivity = 91% [95% CI, 82%-97%] and specificity
Solomon et al., 2016). This marker, referred to as the “central vein sign,” = 96% [95% CI, 88%-100%]) (Castellaro et al., 2020). However,
has gained attention in recent years, as it could help to differentiate MS applying percentage-based criteria requires manual exclusion of lesions
from mimicking diseases (Sati et al., 2016; Sinnecker et al., 2019; Ciotti that are confluent or have multiple or eccentric veins, and performing
et al., 2022; Sparacia et al., 2018; Tranfa et al., 2022). Small cerebral the CVS evaluation on all the remaining lesions present in patients’
veins can be detected with susceptibility-based MRI sequences, taking brains, which is a time-consuming process difficult to accomplish in
advantage of the magnetic properties of venous blood that is rich in clinical practice.
deoxyhemoglobin (Haacke et al., 2009; Mittal et al., 2009). The CVS can Paramagnetic rim lesions (PRL) - Recent pathology studies have

Table 1
Summary of the methods proposed for the automated or semi-automated analysis of cortical lesions, the central vein sign, and paramagnetic rim lesions. The task is
abbreviated as follows: segmentation (S), classification (C). If not specified, all sequences were 3D. Other abbreviations: k-nearest neighbors algorithm (K-NN),
convolutional neural network (CNN), partial volume (PV).
Biomarker Authors (year) Method Task MRI sequences (magnetic field Dataset size Code
strength) (n. of sites) available

Cortical Tardif,C. L., et al. (Tardif et al., 2010) Laminar profile shape analysis S Quantitative high-resolution 1 post mortem brain No
lesions (2010) scan (3 T) scan (1)
Fartaria, M.J., et al. (Fartaria et al., K-NN S FLAIR, MPRAGE, MP2RAGE, 39 MS patients (1) No
2016) (2016) DIR (3 T)
Fartaria, M.J., et al. (Fartaria et al., K-NN with partial volume S FLAIR, MPRAGE, MP2RAGE, 39 MS patients (1) No
2017) (2017) constraints DIR (3 T)
Fartaria, M.J., et al. (Fartaria et al., PV estimation and topological S MP2RAGE (7 T) 25 MS patients (2) No
2019)(2019) constraints

La Rosa, F., et al. (La Rosa et al., 2020) CNN S FLAIR, MP2RAGE (3T) 90 MS patients (2) Yesb
(2020)
La Rosa, F., et al. (La Rosa et al., 2020) CNN S and MP2RAGE, 2D T2*-w GRE, 60 MS patients (1) Yesc
(2020) C T2*w 3D-EPI (7T)
La Rosa, F., et al CNN S MP2RAGE, 2D T2*-w GRE, 80 MS patients (2) Yesc
(2022) (La Rosa et al., 2022) and T2*w 3D-EPI (7T)
C
Paramagnetic rim Barquero, G., et al. (Barquero et al., CNN C FLAIR, T2*w 3D-EPI (3T) 124 MS patients (2) No
lesions 2020) (2020)
Lou, C., et al. (Lou et al., 2021) (2021) Random forest classifier C FLAIR, T1-w, T2*w 3D-EPI (3T) 20 MS patients (1) Yesd
Zhang, H. et al. (2022) (Zhang et al., Residual network and C FLAIR, QSM (3T) 172 MS patients (1) No
2022) radiomic features
Central vein sign Maggi, P., Fartaria, MJ, et al. (Maggi CNN C T2*w 3D-EPI, FLAIR (3T) 42 MS patients, 33 mimics, No
et al., 2020) (2020) 5 others (3)
Dworkin, J. D., et al. (Dworkin et al., Probabilistic method C T2*w 3D-EPI, FLAIR (3T) 16 MS patients, 15 MS Yese
2018) mimics (1)
(2018)
b
https://github.com/FrancescoLR/MS-lesion-segmentation
c
https://github.com/Medical-Image-Analysis-Laboratory/CLaiMS
d
https://github.com/carolynlou/prlr
e
https://github.com/jdwor/cvs

3
F. La Rosa et al. NeuroImage: Clinical 36 (2022) 103205

demonstrated that about 30% of chronic demyelinated lesions are as an outcome measure in clinical trials designed to halt the chronic
pathologically characterized by perilesional accumulation of iron-laden inflammation at the lesion edge. In addition to their prognostic role, PRL
microglia and macrophages, showing evidence of smouldering demye­ appear specific to MS, as they have been rarely detected in patients with
lination and axonal loss around an inactive hypocellular core (see Fig. 4) other neurological conditions (52% of MS vs 7% of non-MS in a multi­
(Frischer et al., 2015; Luchetti et al., 2018). This type of MS lesion has center study of 438 individuals) (Maggi et al., 2020). PRL have the
been defined as “chronic active/smouldering lesions”. Due to their pe­ promise of becoming a clinically relevant biomarker to both improve MS
ripheral paramagnetic iron rim, these lesions can be depicted using in- diagnosis and monitor treatment efficacy over time.
vivo susceptibility-based MRI techniques (T2*-weighted magnitude, Overall, there are not yet imaging guidelines for the visual detection
phase images, and quantitative susceptibility mapping, QSM) at both 3 T of PRL which requires specific training and remains challenging and
and 7 T (Absinta et al., 2018; Absinta et al., 2016), and are therefore time-consuming. The development of ML-based approaches, described
termed “paramagnetic rim lesions” (PRL). in the next section, may help alleviate these issues and facilitate PRL
Direct comparison among different MRI sequences and post­ assessment.
processing techniques for PRL detection is still limited. A recent study Slowly evolving/expanding lesions (SELs) – A different computa­
(Huang et al., 2022) has compared QSM and high-pass-filtered (HPF) tional approach, designed to detect in vivo longitudinal volumetric
phase imaging for identifying PRL. Of 2062 MS lesions detected in 80 lesional changes not associated with gadolinium enhancement, iden­
patients, 9.1% were identified as PRL in both QSM and HPF phase, 9.8% tifies the so-called “slowly evolving/expanding lesions” or SELs. Linear
were PRL only in HPF phase, and the rest were rim negative. QSM- and radial lesion expansion is computed as a function of the Jacobian
identified PRL showed stronger association with clinical disability determinant of the non-linear deformation field between baseline and
compared to those detected by HPF phase imaging. follow up scans (linearity assessment requires a minimum of 3 scans)
Overall, in vivo studies have shown that about 50% of relapsing and (Elliott et al., 2019). Advantages of this approach relate to the use of
about 60% of progressive MS patients have at least one PRL (Absinta retrospective conventional T1-weighted and T2-weighted scans. re-
et al., 2019; Maggi et al., 2020). Of clinical relevance, PRL accrual has analysis of the ORATORIOa clinical trial found reduced rate of T1-
been recently linked to a more aggressive disease course and disability SELs expansion in progressive patients treated with ocrelizumab vs
accumulation at a younger age and/or shorter disease duration (Absinta placebo (Elliott et al., 2019). A recent study showed that SELs are in­
et al., 2019). Reasons for such association directly rely on a few typical dependent predictors of EDSS worsening after a median follow up of 9
features of these lesions: PRL are destructive (Absinta et al., 2016; Kolb years (Preziosa et al., 2022). The neuropathological correlate of SELs is
et al., 2021), they do not remyelinate (Absinta et al., 2016), and they can currently not yet determined and preliminary data showed only modest
expand over time, (Absinta et al., 2019) demyelinating the surrounding correlation with PRL (Elliott et al., 2021).
tissue and injuring axons, as corroborated by the elevation of serum Overall, CL, PRL, and CVS have the potential to considerably

Fig. 1. Representative examples of the three main types of CL. From left to right: 3 T MP2RAGE (0.75 mm isometric), 7 T MP2RAGE (0.5 mm isometric), 7 T T2*-EPI
(0.5 mm isometric) and 7 T T2*-GRE (0.5 mm isometric). CL, including leukocortical, intracortical, and subpial subtypes, are seen better at 7 T due to higher signal-
to-noise ratios, allowing higher resolution scans, and increased susceptibility effects. The 7 T MP2RAGE image shown was obtained as the average of 4 acquisitions.

neurofilament light chain in patients with PRL who are not forming new
white matter lesions (Maggi et al., 2021). The recent discovery that the
paramagnetic rim can significantly shrink or disappear (Absinta et al., a
A phase 3, randomized, parallel-group, double-blind, placebo-controlled
2021; Dal-Bianco et al., 2021) holds promise regarding its potential use trial.

4
F. La Rosa et al. NeuroImage: Clinical 36 (2022) 103205

Fig. 2. examples of CL seen in different MRI contrasts at 3 T. From left to right: MP2RAGE, DIR, PSIR, IR-SWIET, T2*. Red arrows point to leukocortical lesions and
blue arrows to subpial lesions. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 3. A central vein running through a lesion visible in the three planes (zoomed-in boxes) in a 3D FLAIR* obtained combining FLAIR and T2*-EPI acquisitions at 3
T. Resampling was applied to the magnified images for visualization purposes. FLAIR, T2*-EPI and FLAIR* are the MRI contrasts that have been used by ML ap­
proaches for CVS detection. Refer to the supplementary material for additional examples of the CVS on different susceptibility-weighted imaging sequences.

improve the specificity of MS diagnosis (Junker et al., 2020; Maggi et al., 2016; Tardif et al., 2010; Barquero et al., 2020; Lou et al., 2021; Maggi
2018; Maggi et al., 2018). Moreover, studies have shown that CL, PRL, et al., 2020; Dworkin et al., 2018). In the next section, we describe the
and SELs can be useful to assess prognosis (Calabrese et al., 2012; challenges these approaches have been facing and how these differ from
Absinta et al., 2016). Their manual assessment, however, particularly the segmentation of WML.
for CL, is both time-consuming and prone to inter-rater variability. As for
conventional WML, some automated or semi-automated methods have
been proposed to accelerate this task (Fartaria et al., 2017; Fartaria 2.1. Added challenges for CL, PRL, and CVS assessment
et al., 2019; La Rosa et al., 2020; La Rosa et al., 2020; Fartaria et al.,
Compared to conventional imaging biomarkers, the visual

5
F. La Rosa et al. NeuroImage: Clinical 36 (2022) 103205

Fig. 4. (A) Representative paramagnetic rim lesion seen on a 3 T T2*-weighted seg-EPI magnitude and unwrapped filtered phase in the three orthogonal planes
(zoomed-in red boxes, the rim is indicated with red arrows). The central vein (yellow arrows) is also clearly visible within the lesion. (B) Representative periven­
tricular MS lesion with a paramagnetic rim. Paramagnetic rims are visible on both unwrapped phase and QSM-reconstructed images (white arrows). (For inter­
pretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

assessment of CL, PRL and CVS present some additional challenges. sensitivity of these acquisition techniques for PRL detection, especially
Imaging and assessment guidelines- The first obstacle is repre­ when implemented at different field strengths.
sented by the lack of consensus guidelines for imaging protocols. These evolving or unclear criteria for CL, the CVS, and PRL, wide
Although efforts have been made to standardize the use of MRI in variety of imaging settings, and lack of clear guidelines for standardized
clinical practice for conventional biomarkers (Wattjes et al., 2021), protocols clearly jeopardize the development and wide use of these
guidelines are still in a preliminary stage for CL, PRL, and the CVS. CL biomarkers and of targeted ML techniques.
were included in the MS diagnostic criteria in 2017 (Thompson et al., Expert assessment - Even for experts, the task of segmenting CL,
2018), but, currently, there is no single gold standard sequence at 3 T for detecting the CVS, or classifying PRL is intrinsically more challenging
their detection in a clinical setting. PSIR, DIR, and MP2RAGE are all than segmenting WML. CL are generally smaller in size and more
recommended by an international consensus (Filippi et al., 2019). affected by partial volume (PV) effects, compared to WML. The cortex is
However, these contrasts remain primarily acquired in research settings convoluted, so lesion shape is not as regular as in WM, and traditional
and are not yet widely used in clinical routine. Moreover, although 7 T methods of radiological evaluation (scrolling through an image stack)
MRI is increasingly used to detect CL, no guidelines have been presented are less effective in this context. The detection of the CVS requires
yet to standardize their imaging sequences and their identification. susceptibility-based MRI and its exclusion criteria need to be carefully
Regarding the CVS, in a 2016 consensus statement, the North considered when performing its assessment (Sati et al., 2016).
American Imaging in MS Cooperative (NAIMS) proposed a standard Susceptibility-based images used to detect PRL present variability in the
radiological definition and suggested specific MRI acquisitions (Sati susceptibility signal and several artifacts, therefore experienced raters
et al., 2016). Following these recommendations, recent studies have are needed. Moreover, as these three biomarkers have been so far mainly
shown that high-resolution T2*w 3D-EPI or FLAIR* improve the detec­ studied in research settings, clinicians do not commonly see them in
tion of the CVS compared to clinical acquisitions (Castellaro et al., 2020; clinical practice and might need specific training and dedicated time to
Suh et al., 2019). Nevertheless, a standardized clinical protocol for CVS perform a proper assessment.
detection is still missing. Among the three aforementioned biomarkers,
PRL is probably the one at the earliest stages. Although recent studies
support the feasibility of its assessment on clinical scans and its utility in 2.2. Machine learning specific challenges
improving the diagnosis and prognosis of MS (Maggi et al., 2020), there
are currently no international guidelines for its definition nor a stan­ From a ML perspective, the automated segmentation or classification
dardized MRI protocol for its analysis. Several different imaging mo­ of CL, PRL, and the CVS faces new challenges as compared to their WML
dalities have been used for the PRL assessment, including phase 3D-EPI, counterparts.
susceptibility weighted imaging (SWI), QSM, and multi-echo T2* GRE at Limited datasets - An additional limitation, particularly for super­
both 3 T and 7 T (Absinta et al., 2018; Absinta et al., 2016). However, vised DL-based approaches, is the scarcity and limited size of datasets in
there is a paucity of studies that have systematically compared the which these biomarkers were manually annotated. For their assessment,
CL, CVS, and PRL all require advanced MRI sequences at high or ultra-

6
F. La Rosa et al. NeuroImage: Clinical 36 (2022) 103205

high magnetic field and experienced raters, and this makes it difficult to are no ML-based approaches to assess SELs yet, the prospect of analyzing
have large multi-site datasets. Although national MS registries exist in this additional biomarker with ML is presented in the Discussion section.
most countries, the data sharing of MRI in MS is still limited and often Overall, many fewer methods have been proposed for the assessment of
includes only conventional sequences (Vrenken et al., 2021). Moreover, CL, PRL, and the CVS compared to WML. In what follows, we briefly
the CVS or the rim-shape in PRL are visible only on a few slices per describe these state-of-the-art techniques by grouping them according to
lesion, reducing, even more, the data available to train a supervised the biomarker they assess. A summary of the main characteristics for
approach. each method is presented in Table 1, and a scheme of the MRI sequences
Inter-rater variability - The lack of standardization for both the used to detect these three biomarkers at both 3 T and 7 T is shown in
definition and imaging of these biomarkers contributes to a modest Fig. 5.
inter-rater variability. Barquero et al. (Barquero et al., 2020) showed
that, in a cohort of 124 MS patients, approximately 38% of PRL needed a
consensus review from two raters classifying PRL independently (Cohen 3.1. Cortical lesions
k of 0.73). Absinta et al. observed similar inter-rater agreement between
three experts at 3 T (Fleiss coefficient of 0.71), with somewhat higher ML-based methods automatically segmenting CL have been explored
intra-rater reliability (Cohen k of 0.77) (Absinta et al., 2018). Similar with both 3 T and 7 T MRI. The first work (Tardif et al., 2010) present in
values were reported at 7 T for the same set of patients, whereas the the literature considered a postmortem MS brain imaged at 3 T with
agreement between 3 T and 7 T annotations was substantial (Cohen k of different sequences (T1, T2, and relative proton density) at high reso­
0.78). In a similar way, the inter-rater agreement was shown to be lution (0.35 mm isotropic) (Tardif et al. (2012)). Tardif et al. (Tardif
moderate for the segmentation of CL (Harrison et al., 2015; Nielsen et al., 2010) proposed to first identify the cortical and white matter
et al., 2012; Geurts et al., 2011) and high, but not perfect, for the CVS surfaces, then extract laminar profiles between the two tissues, and
(Cohen k of 0.9) (Maggi et al., 2020; Kau et al., 2013). Imaging quality finally apply a k-means classifier to the profile intensity and shape fea­
and motion artifacts are other factors to consider as they can result in tures to parcellate the cortex and detect lesions. Although showing
inconspicuity of all three biomarkers and, therefore, contribute to poor promising results on one postmortem MS brain, this method was never
inter-rater agreement. Overall, the inter-rater variability represents an validated with larger cohorts nor in-vivo data. A few years later, Fartaria
additional challenge for the development of automated approaches, as et al. (Fartaria et al., 2016) proposed the first automated method for the
there might be large inconsistencies in the annotations of the training or segmentation of both WM and cortical lesions. In their study, they
testing set due to different raters performing the manual assessment. analyzed a cohort of 39 early-stage MS patients and considered both
conventional (FLAIR, MPRAGE) and advanced (MP2RAGE, DIR) MRI
3. Methods sequences at 3 T. In a nutshell, their method consisted of co-registering
the different MRI contrasts, leveraging prior tissue probability maps
Despite the recent discovery of the CVS and PRL and the above- from existing brain atlases of healthy subjects, and finally classifying
mentioned challenges, a few groups have already attempted to sup­ each voxel either as being a lesion or healthy tissue with a k-nearest
port their analysis with automated or semi-automated ML methods. To neighbor (k-NN) algorithm. Additionally, as post-processing, all lesions
these two novel biomarkers, we add also CL, which, although studied for smaller than 3.6 µL were discarded, and a region-growing algorithm was
several years, have only recently been assessed automatically. As there applied to improve the lesion delineation. Results were promising,
showing a CL detection rate of 62% when advanced imaging (FLAIR,

Fig. 5. Scheme showing the main MRI sequences used for detecting each biomarker at both 3 T and 7 T.

7
F. La Rosa et al. NeuroImage: Clinical 36 (2022) 103205

MP2RAGE, and DIR) was included. An extension of this segmentation

framework with a Bayesian partial volume (PV) estimation method was
presented by the same authors (Fartaria et al., 2017). They argued that
CL, being generally small and located at the interface between WM and
GM, suffer from strong PV effects. The addition of this PV model indeed
improved the delineation of CL in terms of both total lesion volume
estimation and dice coefficient (Fartaria et al., 2017).
The same research group also proposed a different segmentation
method for WML and CL using only 7 T MP2RAGE images (called
MSLAST: Multiple Sclerosis Lesion Analysis at Seven Tesla) (Fartaria
et al., 2019). MSLAST computes tissue concentration maps with a PV
algorithm and unifies them based on topological constraints. A
connected-components analysis is then performed on gray matter and
cerebrospinal fluid maps, and small components are classified as MS
lesions. This method was evaluated with 25 MS patients’ scans from two
research centers and reached a 58% patient-wise CL detection rate
(when 6 μL was considered as minimum lesion volume) with a false
positive rate of 40%. Moreover, it showed promising scan-rescan
repeatability within the same session, with a mean total lesion volume
difference (WML and CL combined) of 0.29 mL (mean total lesion vol­
ume 5.52 mL), vs 0.13 mL for the manual segmentations. More recently,
DL-based approaches have been presented as well (La Rosa et al., 2020;
La Rosa et al., 2020). In the first study, La Rosa et al. proposed a
framework for the automated segmentation of WML and CL at 3 T using
FLAIR and MP2RAGE (La Rosa et al., 2020). Their method extracts 3D
patches of 88x88x88 voxels from the two MRI contrasts and feeds them
Fig. 6. Three representative axial slices from one MS patient showing the CL
to a convolutional neural network (CNN). The CNN, inspired by the U-
segmentation results of an automated CL segmentation method (La Rosa et al.,
Net, has an encoder and decoder path, each one with three resolution
2020). 7 T MP2RAGE (left column) and CL mask (right column) showing true
levels. Evaluated on two datasets acquired in different centers, for a total positives (green), false negatives (red), and false positives (blue) of the auto­
of 90 MS patients, the framework showed competitive performance, mated approach with respect to the CL manual segmentation. (For interpreta­
with a CL detection rate of 76% and a false positive rate of 29%. tion of the references to colour in this figure legend, the reader is referred to the
In a second study, the same group proposed a similar approach, this web version of this article.)
time tailored exclusively for the detection of CL using multi-contrast 7 T
MRI (La Rosa et al., 2020). The contrasts considered were MP2RAGE, (Valcarcel et al., 2018; Valcarcel et al., 2018) and obtain a map of the
T2*-weighted GRE, and T2*-w 3D-EPI. A cohort of 60 patients was veins by applying the vesselness filter to a T2*w 3D EPI image. Confluent
analyzed with a total of over 2000CL manually segmented by two ex­ lesions are then separated, and lesion centers are detected by textural
perts. The CNN architecture proposed was similar to the one just analysis (Dworkin et al., 2019). Periventricular lesions are discarded as
described, but with a modified output. In addition to the CL segmenta­ suggested by consensus guidelines (Sati et al., 2016), and a permutation
tion, the CNN provided a classification into two types (leukocortical and algorithm is applied to verify whether veins occur at the lesions’ centers
intracortical/subpial lesions) and a separate branch with a simple tissue more often than would be expected due to random chance. Finally, to
segmentation in WM/GM. CL were correctly classified into the two types account for scan motion, the single lesion CVS + probabilities are
by the network with an accuracy of 86%. Setting a minimum lesion size weighted by the noise in their T2*-w 3D-EPI intensities and averaging
of 0.75 μL, it achieved a CL detection rate of 67% with, however, a quite across the total number of lesions, a patient-wise CVS value is obtained.
high false positive rate of 42% (see Fig. 6). Importantly, about 24% of This method was evaluated on a cohort of 31 adults, of whom 16 had
these false positives were retrospectively judged as CL or possible CL by MS. When considering a 40% cutoff rule, the method yielded a sensi­
an expert (La Rosa et al., 2020). tivity of 0.94 and a specificity of 0.67 on a patient-wise classification
In a following publication (La Rosa et al., 2022), this method was level. The performance of the method on a lesion-wise level was not
further improved and evaluated on a multi-site dataset. Its main modi­ assessed. Although still far from experts’ performance, this was a first
fications included an added resolution level in the CNN architecture, a attempt to automatize the CVS assessment and encouraged further
larger 3D patch input size of 96x96x96 voxels, and the use of the focal improvements.
loss for training. Finally, a domain adaptation approach was applied to Maggi, Fartaria et al. (Maggi et al., 2020) introduced an optimized
verify the performance on external datasets. On 20 MRI scans of patients CNN for the automated CVS assessment, called CVSnet. CVSnet is
imaged in a different center, this method achieved superior performance inspired by the VGGnet (Simonyan and Zisserman, 2015) but composed
(CL detection rate of 71%) compared to MSLAST (48%) when setting a of only three convolutional layers followed by rectified linear unit
minimum lesion size of 6 μL. (ReLU) activations. Dropout was applied in each layer, and then two
fully-connected layers of size 32 and 2, respectively, were added to
3.2. The central vein sign provide the output. The authors selected 3D patches of size 21x21x21
voxels as input for the network, where each patch was centered on an MS
As of today, two automated ML methods for the classification of MS lesion and FLAIR* was the only MRI contrast used. Moreover, an
lesions as CVS+ (MS lesions showing the presence of the CVS) or CVS- ensemble of 10 networks with the same architecture was trained and the
(MS lesions without the CVS) have been proposed in the literature probability outputs were averaged to provide the final prediction. This
(Maggi et al., 2020; Dworkin et al., 2018). Both approaches were study considered a cohort of 80 patients imaged at three different sites,
developed and evaluated only with 3 T MRI. Dworkin et al. (Dworkin of whom 42 had MS, 35 an MS-mimic, and 5 an unknown diagnosis. On
et al., 2018) proposed a probabilistic method based on the Frangi ves­ the test set, CVSnet reached a lesion-wise sensitivity, specificity, and
selness filter (Frangi et al., 1496). They first perform an automated WML accuracy of 0.83, 0.75, and 0.79, respectively. On a patient-wise level,
segmentation using T1 and FLAIR 3D MRI volumes acquired at 3 T

8
F. La Rosa et al. NeuroImage: Clinical 36 (2022) 103205

using a 50% cut-off, CVSnet achieved a sensitivity, specificity, and ac­ convolutional features from QSM and FLAIR images, whereas the second
curacy of 0.89, 0.92, and 0.90, respectively, outperforming the vessel­ consists in a fully-connected network that processes previously obtained
ness filter (Frangi et al., 1496) and approaching expert performance. radiomic features. Convolutional and radiomic features are concate­
However, as argued by the authors, these results are not directly com­ nated and a minority oversampling network is used to alleviate the issue
parable with those of Dworkin at al. (Dworkin et al., 2018), as the of class imbalance. Finally, a probability of being a PRL is assigned to
CVSnet considered different exclusion criteria to pre/select the lesions, each lesion. QSM-RimNet was evaluated with a stratified 5-folds cross-
and the initial lesion segmentation was performed manually. validation over 172 MS patients with a total of 177 PRL. Compared to
RimNet and the automated approach of Lou et al., it outperformed both
3.3. Paramagnetic rim lesions methods achieving a lesion-wise sensitivity and specificity of 0.68 and
0.99, respectively, although the differences were not statistically sig­
To our knowledge, only three methods have been proposed so far for nificant. Ablation studies showed that fusing convolutional and radio­
the detection of rim-like features and classification of PRL (Barquero mic features improves the PRL identification (Zhang et al., 2022). Of
et al., 2020; Lou et al., 2021; Zhang et al., 2022). All three methods note, QSM-Rimnet is not fully-automated as during training and evalu­
considered 3 T MRI sequences, whereas 7 T imaging has not yet been ation it relies on manual corrections by experts of both PRL and
explored for the automated assessment of PRL. Barquero et al. (2020) confluent lesions. Similarly to RimNet, this strong limitation currently
introduced a DL-based approach (called RimNet) for the semi- prevents its wider deployment and applicability.
automated classification of PRL, which considered 3D FLAIR and Overall, two methods have tackled the PRL detection problem
T2*w 3D-EPI and phase 3D-EPI images. RimNet’s architecture is inspired considering mainly the T2*-w 3D-EPI sequence and one method has
by the VGGnet (Simonyan and Zisserman, 2015) and composed of two focused on the QSM. Thus, none of the three frameworks has investi­
parallel CNN (one for either FLAIR or T2*w 3D-EPI image and one for gated the effect of differences in SWI and QSM processing on ML-based
the phase 3D-EPI image), where each CNN is made of three convolu­ tools performance and this important aspect should be explored in
tional layers followed by a max-pooling operation. 3D patches of size future studies.
28x28x28 (centered around each MS lesion) are fed to each branch, and
both high-level and low-level feature maps are concatenated. An auto­ 4. Discussion
mated lesion segmentation based on FLAIR and MPRAGE/MP2RAGE (La
Rosa et al., 2020; La Rosa et al., 2019) was modified by an expert to split The methods described in the present review tackle challenging and
confluent lesions. The performance of RimNet was assessed on a cohort clinically relevant problems. Automated and reliable solutions for
of 124 adults with MS who underwent 3 T MRI at two different sites with detecting, classifying, and segmenting CL, PRL, and CVS are needed to
two scanners from the same vendor. Two experts annotated PRL inde­ improve the standardization of these biomarkers and facilitate their
pendently and reached consensus in a joint session (462 PRL in total). assessment in clinical routine. As of today, however, these methods are
The proposed multimodal approach based on FLAIR and phase 3D-EPI still in an early stage and are slightly less sensitive than WML segmen­
image achieves lesion-wise sensitivity and specificity of 0.70 and 0.95, tation approaches.
respectively. When considering a previously identified clinical threshold Nevertheless, such tools would provide obvious advantages, either as
of 4 PRL (Oladosu et al., 2021) for classifying patients as “chronic stand-alone or adjunctive approaches as all three biomarkers are diffi­
active” and “non-chronic active”, RimNet reaches an accuracy of 0.90 cult and time-consuming to analyze using conventional radiological
and an F1-score of 0.84. These values are within 5% of the single ex­ workflows. In these particular cases, manual reading is so involved that
perts’ metrics, suggesting that RimNet could be a valuable tool in sup­ automated methods might actually boost the biomarkers’ widespread
porting the PRL analysis. The main drawback of RimNet, however, is adoption. First, they can substantially reduce analysis time, as compared
that the method is not fully automated, as confluent lesions were split to a manual rating. Maggi, Fartaria et al., for instance, showed that
manually by an expert. CVSNet was 600-fold faster on the test set compared to the manual
Lou et al. (Lou et al., 2021), on the other hand, proposed a fully assessment (4 s vs 40 min) when considering a 50% CVS + lesions
automated ML method for PRL assessment. They considered a cohort of criteria to distinguish MS from MS mimics (Maggi et al., 2020). A lower
20 subjects with MS imaged with 3D FLAIR, 3D MPRAGE, and T2*-w time gain, however, would be expected if CVS + lesion-count criteria,
3D-EPI and phase 3D-EPI images. One neurologist inspected the T2* such as the 3-lesion and 6-lesion, were to be considered. Reduced
magnitude and unwrapped phase images and annotated PRL (113 PRL analysis time can be predicted also for PRL and CL assessment. In La
over the entire cohort). The automated pipeline, after some pre- Rosa et al. (La Rosa et al., 2020), for instance, the automated CL seg­
processing steps that included lesion segmentation (Valcarcel et al., mentation of one subject is computed on average in 20 s. Although a
2018; Valcarcel et al., 2018), lesion center detection (Dworkin et al., direct comparison with the manual labeling was not reported, seg­
2019), and lesion labeling, consisted of extracting 44 different lesion- menting CL manually is known to be a much more time-consuming
wise radiomic features. A random forest classifier was then fitted on process. A second main advantage of automated ML methods is their
these features, and its ability to classify PRL was evaluated on a test set ability to base their decision on 3D multi-contrast MRI analyzed
of 4 patients. Sensitivity and specificity of 0.75 and 0.81, respectively, simultaneously. This stands in contrast to expert reviews, which typi­
were achieved. Although fully automated, this study has three limita­ cally involve comparison of 2D slices across several contrast mecha­
tions. First, the extremely small testing dataset (4 patients only with 47 nisms in a variety of planes and are thus inherently limited in the
PRL), annotated by a single expert, does not guarantee the generaliza­ amount of information that can be readily gleaned.
tion of the proposed method. Second, all patients analyzed had at least
one PRL, and this might add a bias to the trained model. Finally, as 4.1. Common trends
acknowledged by the authors, about 65% of misclassified lesions were
confluent, highlighting the need for a better solution to address these Some common trends can be observed in most of the proposed
lesions. pipelines. The large majority of the methods are supervised, relying on
Inspired by these two previous works, Zhang et al. introduced QSM- expert annotations. Regarding the DL-based approaches, they all used
RimNet (Zhang et al., 2022), a QSM-based approach that combines a patch-based 3D CNN, exploiting the 3D intrinsic information, and often
two-branch feature extraction network and a synthetic minority over­ considered more than one MRI contrast simultaneously. In addition, a
sampling technique. QSM-RimNet receives as input 3D patches of size shared tendency consists of the use of relatively shallow architectures,
32x32x16 voxels where a masking out of non-lesional area is applied. with a limited number of trainable parameters, due to the lack of large
One branch of the network employs residual blocks to extract datasets (La Rosa et al., 2020; La Rosa et al., 2020; Barquero et al., 2020;

9
F. La Rosa et al. NeuroImage: Clinical 36 (2022) 103205

Maggi et al., 2020). Combining this with extensive data augmentation were excluded based on the NAIMS criteria (Sati et al., 2016), whereas
techniques can help when datasets are small and unbalanced. Alterna­ in the pipeline proposed by Dworkin et al. (Dworkin et al., 2018), scans
tively, other groups have tackled the problem of overfitting by proposing affected by noise were discarded following a manual rating. Similarly,
approaches based on classical ML techniques, such as k-NN (Fartaria RimNet (Barquero et al., 2020) exploits lesion masks where confluent
et al., 2017; Fartaria et al., 2016) or random forest classifier (Commo­ lesions have been previously split into single units by an expert. In
wick et al., 2018). In these studies, either intensity-based, radiomic, or contrast, all methods described to date for CL segmentation or detection
probabilistic features are extracted and then fed to the respective clas­ are fully automated (Fartaria et al., 2017; La Rosa et al., 2020; La Rosa
sifier. Overall, their current performance is inferior compared to their et al., 2020; Fartaria et al., 2016). Another persistent issue in the auto­
DL-based counterparts. mated analysis of the CVS and PRL is the presence of confluent lesions.
In addition, some common pre-processing steps can also be identi­ Large, periventricular white matter lesions which include several single
fied. First, some methods use intensity normalization techniques, either units pose additional challenges as the current methods classify each
based on entire 3D volumes (Lou et al., 2021; Dworkin et al., 2018; lesion singularly (Lou et al., 2021; Dworkin et al., 2018), and some of
Fartaria et al., 2017; Fartaria et al., 2019; La Rosa et al., 2020; La Rosa them extract 3D patches centered on the lesion of interest (Barquero
et al., 2020) or on single patches (Barquero et al., 2020; Maggi et al., et al., 2020; Maggi et al., 2020). In RimNet (Barquero et al., 2020), for
2020). Second, the approaches using multiple MRI contrasts always instance, an expert manually split confluent lesions, whereas Lou et al.
register all images to the same space (Lou et al., 2021; Fartaria et al., observed a consistent drop in performance in PRL classification in the
2017; Fartaria et al., 2019; La Rosa et al., 2020; La Rosa et al., 2020). presence of confluent lesions (Lou et al., 2021). Although methods to
Registration errors might affect the methods’ performance. Finally, a automatically split confluent lesions have been proposed (Dworkin
shared pre-processing step in all approaches for the CVS or PRL assess­ et al., 2019; Zhang et al., 2021), further developments are needed in
ment is the prior WML segmentation, obtained either manually (Maggi order to properly apply these in the presence of the CVS or PRL.
et al., 2020) or with an automated tool (Barquero et al., 2020; Lou et al., Finally, for every automated tool the regulatory environment re­
2021; Dworkin et al., 2018). In both cases, this can be a source of error mains a critical barrier, as up to date less than 90 AI/ML-based medical
that negatively affects the subsequent biomarkers’ classification devices or algorithms have been approved by the US Food & Drugs
accuracy. Administration (FDA). This challenge, however, is not unique to the
three biomarkers considered (Pinto et al., 2020) but shared also by
4.2. Current limitations automated approaches segmenting WML or estimating brain atrophy.

Currently, a major limitation hinders the deployment of the above- 4.3. Future research avenues
described methods to the clinic: the methods proposed were trained
and evaluated on small datasets acquired from one or at most two Standardization of the biomarkers’ assessment- The first two
centers. Moreover, the MRI protocols used were often similar and not necessary steps toward the improvement of the above-referred ap­
representative of the current diversity of images acquired in the clinics, proaches are the validation of the biomarkers’ specific criteria and
including different processing, scans affected by noise and artifacts or standardization of the relative MRI protocols. CL have been recently
protocols missing certain modalities. Therefore, the automated ML included in the MS diagnostic criteria (Thompson et al., 2018), however,
methods’ robustness on larger datasets and different scanners, especially a consensus on imaging and on their definition is still missing. In a
from multiple vendors, remains to be proven. This limitation is similar way, PRL urgently need a consensus definition and standardized
emphasized by the current lack of standardized acquisition protocols clinical protocols, whereas the initial criteria proposed for the CVS (Sati
which increases the diversity of the MRI sequences considered for the et al., 2016) need to be updated in light of the latest studies. This would
same biomarkers. This also represents a major hurdle for potential clarify the automated methods’ goals, which so far have been extremely
regulatory approval of such methods. As regulatory approval is neces­ dependent on specific expert labeling of each dataset or on the specific
sary for widespread adoption in the clinics, which is, in turn, the pre­ criteria adopted.
requisite for the availability of large datasets, this is currently a circular Standardization and extensive validation of the automated
dependency issue. methods - Currently, it is difficult to compare the performance of
In addition, the achieved performance levels of the automated ML automated ML methods considering different criteria (such as the min­
methods are still inferior compared to the human experts. Considering imum lesion size) and being evaluated on private datasets. In the future,
the high inter-rater variability and the limited amount of data available, the generalization of the proposed methods should be validated on large,
there is also a considerable risk of having methods that perform well on multi-site datasets with standardized metrics. For this purpose, we urge
data annotated by a single expert and not as well with annotations from research groups to organize grand challenges and release publicly
other raters. To mitigate this issue, several methods have already available datasets with manual annotations of CL, PRL, and CVS. As
considered consensus annotations from two or more experts (La Rosa already proved for several other tasks in medical imaging (Antonelli
et al., 2020; Barquero et al., 2020; Maggi et al., 2020). Regarding CL, no et al., 2021), including for WML segmentation (Carass et al., 2017;
automated method presented in the literature was compared, on the Commowick et al., 2018), such open data initiatives boost on the one
same dataset, with the experts’ inter-rater variability, thus a proper hand the development of state-of-the-art methods, and on the other
evaluation is not possible. With respect to CVS, Maggi, Fartaria et al. hand, help set benchmarks for a fair assessment. Only 5 of the 12
(Maggi et al., 2020) compared the performance of CVSnet with the methods covered in this review are publicly available. In order to extend
consensus of two experts. Following the “50% rule,” CVSnet achieved on their usage and foster a culture of open science, research groups should
the testing set a classification accuracy of 79%, whereas the experts make their code publicly available and possibly provide Docker (Docker,
reached 100% accuracy in differentiating MS and mimic diseases. In a 2014)/Singularity (Kurtzer et al., 2017) images to facilitate their eval­
similar way, Barquero et al. (Barquero et al., 2020) compared RimNet’s uation. Moreover, as successfully done for WML segmentation (Valverde
performance with those of two experts in classifying PRL. In a lesion- et al., 2019), domain-adaptation techniques should also be explored in
wise analysis, RimNet achieved a sensitivity of 71% and a negative order to improve robustness of the automated ML methods to noise,
predictive value of 96%, approaching the experts, who reached 78% and artifacts, and different protocols. So far, all three biomarkers have been
98%, respectively. primarily studied at 3 T and 7 T, and therefore robust methods able to
Another main limitation is represented by the fact that some methods work with images acquired at both magnetic field strengths would be
presented are not fully automated. CVSnet (Maggi et al., 2020), for very valuable. Machine learning algorithms could exploit 7 T enhanced
instance, used manually annotated MS lesion masks in which lesions spatial resolution and tissue contrast by domain adaptation techniques

10
F. La Rosa et al. NeuroImage: Clinical 36 (2022) 103205

to improve their performance on 3 T imaging, which will continue to be Explainable AI (XAI) methods are needed as to on one side provide
the main tool for clinicians as well as for clinical research and trials for uncertainty estimates regarding the output provided and on the other
the foreseeable future. Although it would be highly desirable to have side transparency on the decisions taken by the DL-models. By
methods that work also at the most accessible field strength of 1.5 T, this explainability, we refer to a set of domain features such as pixels of an
seems currently unlikely as the sensitivity to these biomarkers is field- image or human-understandable high-level attributes that contribute to
dependent. the output decision of the model and its internal working. To our
Transfer learning - Considering the scarcity of large, annotated knowledge, there are only two groups that have investigated XAI in MS.
datasets, an additional strategy that should be explored consists of Eitel et al. (Eitel et al., 2019) explored explainability to reveal relevant
transfer learning. Sharing of neural network weights between research voxel-wise locations that a trained CNN uses for distinguishing between
groups could foster interdisciplinary applicability of CNN trained on a normal and MS brain MRI. They found that diagnostic success relied on
relatively large datasets towards different purposes, such as CL, PRL, and the appearance of both lesions and non-lesional tissue (thalamus). Nair
CVS, by fine-tuning the trained models in smaller datasets. Potential et al.Nair et al. (2020) studied the uncertainty of DL-based lesion seg­
advantages would include a shorter training time and robust feature mentation to quantify the AI model reliability. Interestingly, their results
extraction across different MRI device manufacturers or different pulse showed that discarding lesions with high estimated uncertainty from the
sequence acquisition parameters (Valverde et al., 2021). output segmentation would improve the performance of the model.
Longitudinal assessment - Another possible research direction is an These two pioneering approaches strengthen the idea that explainability
expansion of the current methods to analyze longitudinal data. To the and uncertainty measures can reliably provide new insights into how DL
best of our knowledge, only one study has tackled the automated lon­ models for MS work and potentially improve them and increase their
gitudinal assessment of CL at 3 T (Fartaria et al., 2019), whereas PRL transparency.
evolution over time has not yet been assessed with automated ap­ Overall, we believe that developing explainable AI tools is crucial in
proaches. CL are known to play a major role in disease progression the ML MS research roadmap and would have an impact at both meth­
(Mainero et al., 2015) and considerable changes in their volume were odological and clinical levels. First, explainable DL in MS would provide
observed in longitudinal studies (Calabrese et al., 2008; Faizy et al., new insights into model decisions and help identify any bias. Second, the
2019). Of similar interest, PRL and slowly-evolving lesions (SELs) vol­ inclusion of uncertainty and explainability will help in increasing the
ume assessment over time is a plausible future clinical measure of confidence of clinicians considering their use, as well as improve the
treatment response (Absinta et al., 2021; Dal-Bianco et al., 2021; Elliott quality of decision making and ultimately the clinical impact. Finally,
et al., 2019; Elliott et al., 2019). Therefore, automated longitudinal they may foster a better understanding of MS progression by generating
assessment of both CL and PRL could be of high relevance. Regarding biologically interpretable measures of inflammation and degeneration.
SELs, longitudinal WML segmentation approaches (Lladó et al., 2012)
could be adapted to track their evolution in a fully-automated way. This 5. Conclusions
would facilitate their assessment as currently, following an automated
cross-sectional WML segmentation, the lesion masks at each timepoint To summarize, automated or semi-automated ML-based approaches
are manually reviewed (Elliott et al., 2019). aiming to segment and classify CL, CVS, and PRL are still in an early
Joint assessment of multiple biomarkers- To date, all the methods stage. Nevertheless, these pioneering methods have the potential to
proposed tackled the assessment of a single lesional biomarker, although provide standardized identification of the biomarkers and facilitate their
in the case of CL some methods consider WML as well (Fartaria et al., large-scale assessment in clinical routines. Automated or semi-
2019; La Rosa et al., 2020; Fartaria et al., 2016). Future work may aim at automated tools could considerably reduce the current amount of time
automatically analyzing multiple biomarkers in a unified framework and effort needed for a manual assessment. To date, however, some
(eg. with the same input images and algorithm) as this would be limitations still hinder a broader adoption of these tools. First, there is a
extremely useful for research purposes or in clinics. Moreover, ML-based general need for consensus criteria and standardized clinical protocols
algorithms have the potential to be useful also for prediction purposes. A for all three aforementioned biomarkers. Further, a major barrier to the
few automated methods based either on MRI (Tousignant et al., 2021; automated methods’ deployment is their lack of validation on multi-
Marzullo et al., 2019; Roca et al., 2020), optical coherence tomography center datasets acquired with different protocols. Future work should
(Montolío et al., 2021), or clinical information (Pinto et al., 2020) have focus on improving the robustness of the automated methods, extending
already been presented to predict MS progression. Specifically to the their framework with longitudinal data, and including interpretable
biomarkers considered in the present review, Treaba et al. have pro­ measures into their decisions. Finally, we encourage research groups to
posed a ML approach for the regression of both CL and PRL, in the same organize grand challenges and release publicly available datasets. This
cohort of patients, with disability progression (Treaba et al., 2021;3(3): would boost the development of new methods and provide benchmarks
fcab134.). In this prospective, longitudinal study, the authors analyzed for a fair and standardized comparison that is currently lacking.
brain scans of 100 MS patients using 7 T susceptibility-sensitive MRI in
which CL and PRL were segmented manually. Although the study had CRediT authorship contribution statement
some limitations, including the fact that the disability progression was
assessed only by the EDSS and only one ML-based method (gradient Francesco La Rosa: Conceptualization, Methodology, Validation,
boosting algorithm, XGBoost) was tested, it showed that 7 T MRI and the Formal analysis, Investigation, Data curation, Writing – original draft,
combination of different biomarkers are promising in predicting MS Writing – review & editing, Visualization. Maxence Wynen: Method­
disability progression. Future studies should aim to combine the auto­ ology, Visualization, Formal analysis, Investigation, Data curation,
mated assessment of multiple biomarkers with clinical information and Writing – original draft, Writing – review & editing. Omar Al-Louzi:
other relevant markers to predict clinical outcomes or treatment effect. Methodology, Validation, Formal analysis, Data curation, Writing –
Explainable AI - As discussed in this paper, ML methods combined original draft, Writing – review & editing. Erin S Beck: Methodology,
with specialized MRI sequences could play a fundamental role in sup­ Validation, Resources, Data curation, Writing – original draft, Writing –
porting the diagnosis of, and prognostication in, MS. However, the review & editing, Visualization. Till Huelnhagen: Resources, Writing –
complexity of DL algorithms hinders their interpretation, which has led original draft, Writing – review & editing, Visualization. Pietro Maggi:
some to consider these methods as “black boxes.” The lack of an obvious Resources, Writing – original draft, Writing – review & editing, Visual­
connection between biology, pathophysiology, and features revealed by ization. Jean-Philippe Thiran: Investigation, Supervision, Project
DL might diminish clinicians’ confidence in these algorithms, again administration, Funding acquisition. Tobias Kober: Investigation,
hindering the adoption of such tools in clinical research and healthcare. Writing – original draft, Writing – review & editing. Russell T

11
F. La Rosa et al. NeuroImage: Clinical 36 (2022) 103205

Shinohara: Validation, Investigation, Writing – original draft, Writing – References

review & editing. Pascal Sati: Validation, Investigation, Writing –
original draft, Writing – review & editing. Daniel S Reich: Validation, Absinta, M., Sati, P., Schindler, M., et al., 2016. Persistent 7-tesla phase rim predicts poor
outcome in new multiple sclerosis patient lesions. J Clin Invest. 126 (7), 2597–2609.
Formal analysis, Investigation, Resources, Writing – original draft, https://doi.org/10.1172/JCI86198.
Writing – review & editing. Cristina Granziera: Validation, Resources, Absinta, M., Sati, P., Fechner, A., Schindler, M.K., Nair, G., Reich, D.S., 2018.
Writing – original draft. Martina Absinta: Validation, Formal analysis, Identification of chronic active multiple sclerosis lesions on 3T MRI. Am J
Neuroradiol. 39 (7), 1233–1238. https://doi.org/10.3174/ajnr.A5660.
Investigation, Resources, Data curation, Writing – original draft, Writing Absinta, M., Sati, P., Masuzzo, F., et al., 2019. Association of chronic active multiple
– review & editing, Visualization. Meritxell Bach Cuadra: Conceptu­ sclerosis lesions with disability in vivo. JAMA Neurol. 76 (12), 1474. https://doi.
alization, Methodology, Validation, Investigation, Resources, Writing – org/10.1001/jamaneurol.2019.2399.
Absinta, M., Maric, D., Gharagozloo, M., et al., 2021. A lymphocyte–microglia–astrocyte
original draft, Writing – review & editing, Supervision, Project admin­ axis in chronic active multiple sclerosis. Nature 597 (7878), 709–714. https://doi.
istration, Funding acquisition. org/10.1038/s41586-021-03892-7.
Antonelli M, Reinke A, Bakas S, et al. The Medical Segmentation Decathlon.
ArXiv210605735 Cs Eess. Published online June 10, 2021. Accessed November 8,
Declaration of Competing Interest 2021. http://arxiv.org/abs/2106.05735.
Barkhof, F., 2002. The clinico-radiological paradox in multiple sclerosis revisited. Curr.
The University Hospital Basel (USB), as the employer of C.G., has Opin. Neurol. 15 (3), 239–245. https://doi.org/10.1097/00019052-200206000-
00003.
received the following fees which were used exclusively for research
Barquero, G., La Rosa, F., Kebiri, H., et al., 2020. RimNet: A deep 3D multimodal MRI
support: (i) advisory board and consultancy fees from Actelion, architecture for paramagnetic rim lesion assessment in multiple sclerosis.
Genzyme-Sanofi, Novartis, GeNeuro and Roche; (ii) speaker feesfrom NeuroImage Clin. 28, 102412 https://doi.org/10.1016/j.nicl.2020.102412.
Genzyme-Sanofi, Novartis, GeNeuro and Roche; (iii) research support Beck ES, Maranzano J, Luciano NJ, et al. Cortical lesion hotspots and association of
subpial lesions with disability in multiple sclerosis. Mult Scler J. Published online
from Siemens, GeNeuro, Roche. M.A. has received consultancy fees from February 10, 2022:135245852110691. doi:10.1177/13524585211069167.
GSK and Sanofi-Genzyme. P.M. has received support from Biogen and Beck, E.S., Gai, N., Filippini, S., Maranzano, J., Nair, G., Reich, D.S., 2020. Inversion
Cliniques universitaires Saint-Luc Fonds de Recherche Clinique. D.S.R. recovery susceptibility weighted imaging with enhanced T2 Weighting at 3 T
improves visualization of subpial cortical multiple sclerosis lesions. Invest. Radiol.
has received research support from Abata, Sanofi-Genzyme, and Vertex. 55 (11), 727–735. https://doi.org/10.1097/RLI.0000000000000698.
The other authors have no known competing financial interests or per­ Calabrese, M., Filippi, M., Rovaris, M., et al., 2008. Morphology and evolution of cortical
sonal relationships that could have appeared to influence the work re­ lesions in multiple sclerosis. A longitudinal MRI study. NeuroImage. 42 (4),
1324–1328. https://doi.org/10.1016/j.neuroimage.2008.06.028.
ported in this paper. Calabrese, M., Filippi, M., Gallo, P., 2010. Cortical lesions in multiple sclerosis. Nat Rev
Neurol. 6 (8), 438–444. https://doi.org/10.1038/nrneurol.2010.93.
Data availability Calabrese, M., Poretto, V., Favaretto, A., et al., 2012. Cortical lesion load associates with
progression of disability in multiple sclerosis. Brain. 135 (10), 2952–2961. https://
doi.org/10.1093/brain/aws246.
No data was used for the research described in the article. Calabrese, M., Romualdi, C., Poretto, V., et al., 2013. The changing clinical course of
multiple sclerosis: A matter of gray matter: Cortical Pathology and Secondary
Progressive MS. Ann. Neurol. 74 (1), 76–83. https://doi.org/10.1002/ana.23882.
Acknowledgements
Campion, T., Smith, R.J.P., Altmann, D.R., et al., 2017. FLAIR* to visualize veins in white
matter lesions: A new tool for the diagnosis of multiple sclerosis? Eur. Radiol. 27
F.L.R. is supported by the Swiss National Foundation (SNF) Postdoc (10), 4257–4263. https://doi.org/10.1007/s00330-017-4822-z.
Mobility Fellowship (P500PB_206833), the European Union’s Horizon Carass, A., Roy, S., Jog, A., et al., 2017. Longitudinal multiple sclerosis lesion
segmentation: Resource and challenge. NeuroImage. 148, 77–102. https://doi.org/
2020 research and innovation program under the Marie Sklodowska- 10.1016/j.neuroimage.2016.12.064.
Curie project TRABIT (agreement No 765148) and the Novartis Foun­ Castellaro, M., Tamanti, A., Pisani, A.I., Pizzini, F.B., Crescenzo, F., Calabrese, M., 2020.
dation for Medical-Biological Research (#21A032). M.W. is supported The Use of the Central Vein Sign in the Diagnosis of Multiple Sclerosis: A Systematic
Review and Meta-analysis. Diagnostics. 10 (12), 1025. https://doi.org/10.3390/
by a Swiss government excellence scholarship (#2021.0087). O.A. is diagnostics10121025.
supported by a National Multiple Sclerosis Society (NMSS) - American Ciotti, J.R., Eby, N.S., Brier, M.R., et al., 2022. Central vein sign and other radiographic
Brain Foundation Clinician Scientist Development Award (FAN- features distinguishing myelin oligodendrocyte glycoprotein antibody disease from
multiple sclerosis and aquaporin-4 antibody-positive neuromyelitis optica. Mult
1807–32163). E.S.B. is supported by a Career Transition Fellowship Scler J. 28 (1), 49–60. https://doi.org/10.1177/13524585211007086.
from the National Multiple Sclerosis Society. P.S., O.A., E.S.B., and D.S. Clarke, M.A., Pareto, D., Pessini-Ferreira, L., et al., 2020. Value of 3T Susceptibility-
R. are supported by the Intramural Research Program of the National Weighted Imaging in the Diagnosis of Multiple Sclerosis. Am J Neuroradiol. 41 (6),
1001–1008. https://doi.org/10.3174/ajnr.A6547.
Institute of Neurological Disorders and Stroke, National Institutes of Commowick O, Cervenansky F, Cotton F, Dojat M. MSSEG-2 challenge proceedings:
Health, Bethesda, Maryland, USA. R.T.S. is partially supported by Multiple sclerosis new lesions segmentation challenge using a data management and
R01NS060910, R01MH112847, R01MH123550, U01NS116776, and processing infrastructure. In: ; 2021:1. Accessed October 28, 2021. https://hal.inria.
fr/hal-03358968.
R01NS112274 from the National Institutes of Health. The content is
Commowick, O., Istace, A., Kain, M., et al., 2018. Objective evaluation of multiple
solely the responsibility of the authors and does not necessarily repre­ sclerosis lesion segmentation using a data management and processing
sent the official views of the National Institutes of Health. M.A. is sup­ infrastructure. Sci. Rep. 8 (1), 13650. https://doi.org/10.1038/s41598-018-31911-
ported by the Conrad N. Hilton Foundation (Marilyn Hilton Bridging 7.
Cortese, R., Magnollay, L., Tur, C., et al., 2018. Value of the central vein sign at 3T to
Award for Physician-Scientists, grant #17313), the International Pro­ differentiate MS from seropositive NMOSD. Neurology. 90 (14), e1183–e1190.
gressive MS alliance, the Roche Foundation for Independent Research, https://doi.org/10.1212/WNL.0000000000005256.
the Cariplo Foundation (grant #1677), and the FRRB Early Career Cortese, R., Collorone, S., Ciccarelli, O., Toosy, A.T., 2019. Advances in brain imaging in
multiple sclerosis. Ther Adv Neurol Disord. 12 https://doi.org/10.1177/
Award (grant#1750327). 1756286419859722.
We acknowledge access to the facilities and expertise of the CIBM Dal-Bianco, A., Grabner, G., Kronnerwetter, C., et al., 2021. Long-term evolution of
Center for Biomedical Imaging, a Swiss research center of excellence multiple sclerosis iron rim lesions in 7 T MRI. Brain. 144 (3), 833–847. https://doi.
org/10.1093/brain/awaa436.
founded and supported by Lausanne University Hospital (CHUV), Uni­ Merkel D. Docker: lightweight Linux containers for consistent development and
versity of Lausanne (UNIL), Ecole Polytechnique fedérale de Lausanne deployment. Linux J. Published online March 1, 2014. Accessed March 15, 2022.
(EPFL), University of Geneva (UNIGE) and Geneva University Hospitals https://dl.acm.org/doi/abs/10.5555/2600239.2600241.
Dworkin, J.D., Sati, P., Solomon, A., et al., 2018. Automated integration of multimodal
(HUG). We thank Thomas Yu for proofreading the manuscript. MRI for the probabilistic detection of the central vein sign in white matter lesions.
Am J Neuroradiol. 39 (10), 1806–1813. https://doi.org/10.3174/ajnr.A5765.
Appendix A. Supplementary data Dworkin, J.D., Linn, K.A., Oguz, I., et al., 2019. An automated statistical technique for
counting distinct multiple sclerosis lesions. Am J Neuroradiol. 39 (4), 626–633.
https://doi.org/10.3174/ajnr.A5556.
Supplementary data to this article can be found online at https://doi. Eitel, F., Soehler, E., Bellmann-Strobl, J., et al., 2019. Uncovering convolutional neural
org/10.1016/j.nicl.2022.103205. network decisions for diagnosing multiple sclerosis on conventional MRI using layer-

12
F. La Rosa et al. NeuroImage: Clinical 36 (2022) 103205

wise relevance propagation. NeuroImage Clin. 24, 102003 https://doi.org/10.1016/ Junker, A., Wozniak, J., Voigt, D., et al., 2020. Extensive subpial cortical demyelination
j.nicl.2019.102003. is specific to multiple sclerosis. Brain Pathol. 30 (3), 641–652. https://doi.org/
Elliott, C., Wolinsky, J.S., Hauser, S.L., et al., 2019. Slowly expanding/evolving lesions as 10.1111/bpa.12813.
a magnetic resonance imaging marker of chronic active multiple sclerosis lesions. Kaisey, M., Solomon, A.J., Luu, M., Giesser, B.S., Sicotte, N.L., 2019. Incidence of
Mult Scler J. 25 (14), 1915–1925. https://doi.org/10.1177/1352458518814117. multiple sclerosis misdiagnosis in referrals to two academic centers. Mult Scler Relat
Elliott, C., Belachew, S., Wolinsky, J.S., et al., 2019. Chronic white matter lesion activity Disord. 30, 51–56. https://doi.org/10.1016/j.msard.2019.01.048.
predicts clinical progression in primary progressive multiple sclerosis. Brain. 142 Kau, T., Taschwer, M., Deutschmann, H., Schönfelder, M., Weber, J.R., Hausegger, K.A.,
(9), 2787–2799. https://doi.org/10.1093/brain/awz212. 2013. The “central vein sign”: is there a place for susceptibility weighted imaging in
Elliott, C., Belachew, S., Fisher, E., et al., 2021. MRI Characteristics of Chronic MS possible multiple sclerosis? Eur. Radiol. 23 (7), 1956–1962. https://doi.org/
Lesions by Phase Rim Detection and/or Slowly Expanding Properties. Neurology. 96 10.1007/s00330-013-2791-4.
(15 Supplement), 4101. Kaur, A., Kaur, L., Singh, A., 2021. State-of-the-Art Segmentation Techniques and Future
Faizy, T.D., Thaler, C., Ceyrowski, T., et al., 2017. Reliability of cortical lesion detection Directions for Multiple Sclerosis Brain Lesions. Arch. Comput. Methods Eng. 28 (3),
on double inversion recovery MRI applying the MAGNIMS-Criteria in multiple 951–977. https://doi.org/10.1007/s11831-020-09403-7.
sclerosis patients within a 16-months period. Kassubek J, ed. PLOS ONE. https://doi. Kidd, D., Barkhof, F., McConnell, R., Algra, P.R., Allen, I.V., Revesz, T., 1999. Cortical
org/10.1371/journal.pone.0172923. lesions in multiple sclerosis. Brain. 122 (1), 17–26. https://doi.org/10.1093/brain/
Faizy, T.D., Broocks, G., Thaler, C., et al., 2019. Development of Cortical Lesion Volumes 122.1.17.
on Double Inversion Recovery MRI in Patients With Relapse-Onset Multiple Kilsdonk, I.D., Jonkman, L.E., Klaver, R., et al., 2016. Increased cortical grey matter
Sclerosis. Front. Neurol. 10 https://doi.org/10.3389/fneur.2019.00133. lesion detection in multiple sclerosis with 7 T MRI: a post-mortem verification study.
Fartaria, M.J., Bonnier, G., Roche, A., et al., 2016. Automated detection of white matter Brain. 139 (5), 1472–1481. https://doi.org/10.1093/brain/aww037.
and cortical lesions in early stages of multiple sclerosis: Automated MS Lesion Kolb, H., Absinta, M., Beck, E.S., et al., 2021. 7T MRI Differentiates Remyelinated from
Segmentation. J. Magn. Reson. Imaging 43 (6), 1445–1454. https://doi.org/ Demyelinated Multiple Sclerosis Lesions. Ann. Neurol. 90 (4), 612–626. https://doi.
10.1002/jmri.25095. org/10.1002/ana.26194.
Fartaria MJ, Roche A, Meuli R, Granziera C, Kober T, Bach Cuadra M. Segmentation of Kurtzer, G.M., Sochat, V., Singularity, B.MW., 2017. Scientific containers for mobility of
Cortical and Subcortical Multiple Sclerosis Lesions Based on Constrained Partial compute. Gursoy A, ed. PLOS ONE. https://doi.org/10.1371/journal.pone.0177459.
Volume Modeling. In: Medical Image Computing and Computer Assisted Intervention − La Rosa F, Fartaria MJ, Kober T, et al. Shallow vs Deep Learning Architectures for White
MICCAI 2017. Vol 10435. Lecture Notes in Computer Science. Springer International Matter Lesion Segmentation in the Early Stages of Multiple Sclerosis. In: Crimi A,
Publishing; 2017:142-149. doi:10.1007/978-3-319-66179-7_17. Bakas S, Kuijf H, Keyvan F, Reyes M, van Walsum T, eds. Brainlesion: Glioma, Multiple
Fartaria, M.J., Kober, T., Granziera, C., Bach, C.M., 2019. Longitudinal analysis of white Sclerosis, Stroke and Traumatic Brain Injuries. Vol 11383. Lecture Notes in Computer
matter and cortical lesions in multiple sclerosis. NeuroImage Clin. 23, 101938 Science. Springer International Publishing; 2019:142-151. doi:10.1007/978-3-030-
https://doi.org/10.1016/j.nicl.2019.101938. 11723-8_14.
Fartaria, M.J., Sati, P., Todea, A., et al., 2019. Automated Detection and Segmentation of La Rosa F, Beck ES, Abdulkadir A, et al. Automated Detection of Cortical Lesions in
Multiple Sclerosis Lesions Using Ultra–High-Field MP2RAGE. Invest. Radiol. 54 (6), Multiple Sclerosis Patients with 7T MRI. In: Medical Image Computing and Computer
356–364. https://doi.org/10.1097/RLI.0000000000000551. Assisted Intervention – MICCAI 2020. Vol 12264. Lecture Notes in Computer Science.
Favaretto, A., Lazzarotto, A., Poggiali, D., et al., 2016. MRI-detectable cortical lesions in Springer International Publishing; 2020:584-593. doi:10.1007/978-3-030-59719-1_
the cerebellum and their clinical relevance in multiple sclerosis. Mult Scler J. 22 (4), 57.
494–501. https://doi.org/10.1177/1352458515594043. La Rosa F, Abdulkadir A, Fartaria MJ, et al. Multiple sclerosis cortical and WM lesion
Filippi, M., Preziosa, P., Banwell, B.L., et al., 2019. Assessment of lesions on magnetic segmentation at 3T MRI: a deep learning method based on FLAIR and MP2RAGE.
resonance imaging in multiple sclerosis: practical guidelines. Brain. 142 (7), NeuroImage Clin. Published online June 2020:102335. doi:10.1016/j.
1858–1875. https://doi.org/10.1093/brain/awz144. nicl.2020.102335.
Frangi AF, Niessen WJ, Vincken KL, Viergever MA. Multiscale vessel enhancement La Rosa F, Beck ES, Maranzano J, et al. Multiple sclerosis cortical lesion detection with
filtering. In: Medical Image Computing and Computer-Assisted Intervention — deep learning at ultra-high-field MRI. NMR Biomed. Published online March 31,
MICCAI’98. Vol 1496. Lecture Notes in Computer Science. Springer Berlin 2022. doi:10.1002/nbm.4730.
Heidelberg; 1998:130-137. doi:10.1007/BFb0056195. Lladó, X., Oliver, A., Cabezas, M., et al., 2012. Segmentation of multiple sclerosis lesions
Frischer, J.M., Weigand, S.D., Guo, Y., et al., 2015. Clinical and pathological insights into in brain MRI: A review of automated approaches. Inf. Sci. 186 (1), 164–185. https://
the dynamic nature of the white matter multiple sclerosis plaque: dynamic nature of doi.org/10.1016/j.ins.2011.10.011.
MS Plaque. Ann. Neurol. 78 (5), 710–721. https://doi.org/10.1002/ana.24497. Lou C, Sati P, Absinta M, et al. Fully Automated Detection of Paramagnetic Rims in
Gaitán, M.I., Yañez, P., Paday Formenti, M.E., et al., 2020. SWAN-Venule: An optimized Multiple Sclerosis Lesions on 3T Susceptibility-Based MR Imaging. NeuroImage Clin.
MRI technique to detect the central vein sign in MS Plaques. Am J Neuroradiol. 41 Published online August 2021:102796. doi:10.1016/j.nicl.2021.102796.
(3), 456–460. https://doi.org/10.3174/ajnr.A6437. Lu, P.J., Yoo, Y., Rahmanzadeh, R., et al., 2021. GAMER MRI: Gated-attention
Geurts JJG, Bo L, Pouwels PJW, Castelijns JA, Polman CH, Barkhof F. Cortical Lesions in mechanism ranking of multi-contrast MRI in brain pathology. NeuroImage Clin. 29,
Multiple Sclerosis: Combined Postmortem MR Imaging and Histopathology. 102522 https://doi.org/10.1016/j.nicl.2020.102522.
Published online 2005:6. Luchetti, S., Fransen, N.L., van Eden, C.G., Ramaglia, V., Mason, M., Huitinga, I., 2018.
Geurts, J.J.G., Roosendaal, S.D., Calabrese, M., et al., 2011. Consensus recommendations Progressive multiple sclerosis patients show substantial lesion activity that correlates
for MS cortical lesion scoring using double inversion recovery MRI. Neurology. 76 with clinical disease severity and sex: a retrospective autopsy cohort analysis. Acta
(5), 418–424. https://doi.org/10.1212/WNL.0b013e31820a0cc4. Neuropathol (Berl). 135 (4), 511–528. https://doi.org/10.1007/s00401-018-1818-y.
Giorgio, A., Stromillo, M.L., Rossi, F., et al., 2011. Cortical lesions in radiologically Madsen, M.A.J., Wiggermann, V., Bramow, S., Christensen, J.R., Sellebjerg, F.,
isolated syndrome. Neurology. 77 (21), 1896–1899. https://doi.org/10.1212/ Siebner, H.R., 2021. Imaging Cortical Multiple Sclerosis Lesions with Ultra-High
WNL.0b013e318238ee9b. Field MRI. Radiology and Imaging. https://doi.org/10.1101/2021.06.25.21259363.
Giorgio, A., Stromillo, M.L., Bartolozzi, M.L., et al., 2014. Relevance of hypointense brain Maggi, P., Absinta, M., Grammatico, M., et al., 2018. Central vein sign differentiates
MRI lesions for long-term worsening of clinical disability in relapsing multiple Multiple Sclerosis from central nervous system inflammatory vasculopathies: Central
sclerosis. Mult Scler J. 20 (2), 214–219. https://doi.org/10.1177/ Vein Sign. Ann. Neurol. 83 (2), 283–294. https://doi.org/10.1002/ana.25146.
1352458513494490. Maggi, P., Fartaria, M.J., Jorge, J., et al., 2020. CVSnet: A machine learning approach for
Granziera, C., Wuerfel, J., Barkhof, F., et al., 2021. Quantitative magnetic resonance automated central vein sign assessment in multiple sclerosis. NMR Biomed. 33 (5)
imaging towards clinical application in multiple sclerosis. Brain. 144 (5), https://doi.org/10.1002/nbm.4283.
1296–1311. https://doi.org/10.1093/brain/awab029. Maggi, P., Sati, P., Nair, G., et al., 2020. Paramagnetic Rim Lesions are Specific to
Haacke, E.M., Mittal, S., Wu, Z., Neelavalli, J., Cheng, Y.C.N., 2009. Susceptibility- Multiple Sclerosis: An International Multicenter 3T MRI Study. Ann. Neurol. 88 (5),
weighted imaging: Technical aspects and clinical applications, Part 1. Am J 1034–1042. https://doi.org/10.1002/ana.25877.
Neuroradiol. 30 (1), 19–30. https://doi.org/10.3174/ajnr.A1400. Maggi, P., Kuhle, J., Schädelin, S., et al., 2021. Chronic White Matter Inflammation and
Hagens, M.H., Burggraaff, J., Kilsdonk, I.D., et al., 2019. Impact of 3 Tesla MRI on Serum Neurofilament Levels in Multiple Sclerosis. Neurology. 97 (6), e543–e553.
interobserver agreement in clinically isolated syndrome: A MAGNIMS multicentre https://doi.org/10.1212/WNL.0000000000012326.
study. Mult Scler J. 25 (3), 352–360. https://doi.org/10.1177/1352458517751647. Mainero, C., Louapre, C., Govindarajan, S.T., et al., 2015. A gradient in cortical
Harrison, D.M., Roy, S., Oh, J., et al., 2015. Association of cortical lesion burden on 7-T pathology in multiple sclerosis by in vivo quantitative 7 T imaging. Brain. 138 (4),
Magnetic Resonance imaging with cognition and disability in multiple sclerosis. 932–945. https://doi.org/10.1093/brain/awv011.
JAMA Neurol. 72 (9), 1004. https://doi.org/10.1001/jamaneurol.2015.1241. Maranzano, J., Dadar, M., Rudko, D.A., et al., 2019. Comparison of Multiple Sclerosis
Hemond, C.C., Bakshi, R., 2018. Magnetic resonance imaging in multiple sclerosis. Cold Cortical Lesion Types Detected by Multicontrast 3T and 7T MRI. Am J Neuroradiol.
Spring Harb Perspect Med. 8 (5), a028969 https://doi.org/10.1101/cshperspect. 40 (7), 1162–1169. https://doi.org/10.3174/ajnr.A6099.
a028969. Martínez-Heras, E., Solana, E., Prados, F., et al., 2020. Characterization of multiple
Huang W, Sweeney EM, Kaunzner UW, Wang Y, Gauthier SA, Nguyen TD. Quantitative sclerosis lesions with distinct clinical correlates through quantitative diffusion MRI.
susceptibility mapping versus phase imaging to identify multiple sclerosis iron rim NeuroImage Clin. 28, 102411 https://doi.org/10.1016/j.nicl.2020.102411.
lesions with demyelination. J Neuroimaging. Published online March 9, 2022: Marzullo, A., Kocevar, G., Stamile, C., 2019. Prediction of multiple sclerosis patient
jon.12987. doi:10.1111/jon.12987. disability from structural connectivity using convolutional neural networks. In, et al.
Ineichen, B.V., Beck, E.S., Piccirelli, M., Reich, D.S., 2021. New Prospects for ultra-high- 41st Annual International Conference of the IEEE Engineering in Medicine and
field magnetic resonance imaging in multiple sclerosis. Invest. Radiol. 56 (11), Biology Society (EMBC). IEEE 2019, 2087–2090. https://doi.org/10.1109/
773–784. https://doi.org/10.1097/RLI.0000000000000804. EMBC.2019.8856845.
Mistry, N., Abdel-Fahim, R., Samaraweera, A., et al., 2016. Imaging central veins in brain
lesions with 3-T T2*-weighted magnetic resonance imaging differentiates multiple

13
F. La Rosa et al. NeuroImage: Clinical 36 (2022) 103205

sclerosis from microangiopathic brain lesions. Mult Scler J. 22 (10), 1289–1296. Solomon, A.J., Naismith, R.T., Cross, A.H., 2019. Misdiagnosis of multiple sclerosis:
https://doi.org/10.1177/1352458515616700. Impact of the 2017 McDonald criteria on clinical practice. Neurology. 92 (1), 26–33.
Mittal, S., Wu, Z., Neelavalli, J., Haacke, E.M., 2009. Susceptibility-Weighted Imaging: https://doi.org/10.1212/WNL.0000000000006583.
Technical Aspects and Clinical Applications, Part 2. Am J Neuroradiol. 30 (2), Sparacia, G., Agnello, F., Gambino, A., Sciortino, M., Midiri, M., 2018. Multiple sclerosis:
232–252. https://doi.org/10.3174/ajnr.A1461. High prevalence of the ‘central vein’ sign in white matter lesions on susceptibility-
Montolío, A., Martín-Gallego, A., Cegoñino, J., et al., 2021. Machine learning in weighted images. Neuroradiol J. 31 (4), 356–361. https://doi.org/10.1177/
diagnosis and disability prediction of multiple sclerosis using optical coherence 1971400918763577.
tomography. Comput. Biol. Med. 133, 104416 https://doi.org/10.1016/j. Suh, C.H., Kim, S.J., Jung, S.C., Choi, C.G., Kim, H.S., 2019. The “Central Vein Sign” on
compbiomed.2021.104416. T2*-weighted Images as a Diagnostic Tool in Multiple Sclerosis: A Systematic Review
Müller, J., La Rosa, F., Beaumont, J., et al., 2022. FLAWS: New Sensitive 3T MRI and Meta-analysis using Individual Patient Data. Sci. Rep. 9 (1), 18188. https://doi.
Contrasts for Cortical Lesion Detection in Multiple Sclerosis. Published online, Invest org/10.1038/s41598-019-54583-3.
Radiol. Tallantyre, E.C., Dixon, J.E., Donaldson, I., et al., 2011. Ultra-high-field imaging
Nair, T., Precup, D., Arnold, D.L., Arbel, T., 2020. Exploring uncertainty measures in distinguishes MS lesions from asymptomatic white matter lesions. Neurology. 76 (6),
deep networks for Multiple sclerosis lesion detection and segmentation. Med. Image 534–539. https://doi.org/10.1212/WNL.0b013e31820b7630.
Anal. 59, 101557 https://doi.org/10.1016/j.media.2019.101557. Tardif, C.L., Bedell, B.J., Eskildsen, S.F., Collins, D.L., Pike, G.B., 2012. Quantitative
Nielsen, A.S., Kinkel, R.P., Tinelli, E., Benner, T., Cohen-Adad, J., Mainero, C., 2012. Magnetic Resonance Imaging of Cortical Multiple Sclerosis Pathology. Mult Scler Int.
Focal cortical lesion detection in multiple sclerosis: 3 tesla DIR versus 7 tesla FLASH- 2012, 1–13. https://doi.org/10.1155/2012/742018.
T2*. J. Magn. Reson. Imaging 35 (3), 537–542. https://doi.org/10.1002/ Tardif CL, Collins DL, Eskildsen SF, Richardson JB, Pike GB. Segmentation of Cortical MS
jmri.22847. Lesions on MRI Using Automated Laminar Profile Shape Analysis. In: Medical Image
Nielsen, A.S., Kinkel, R.P., Madigan, N., Tinelli, E., Benner, T., Mainero, C., 2013. Computing and Computer-Assisted Intervention – MICCAI 2010. Vol 6363. Lecture
Contribution of cortical lesion subtypes at 7T MRI to physical and cognitive Notes in Computer Science. Springer Berlin Heidelberg; 2010:181-188. doi:10.1007/
performance in MS. Neurology. 81 (7), 641–649. https://doi.org/10.1212/ 978-3-642-15711-0_23.
WNL.0b013e3182a08ce8. Thaler, C., Kyselyova, A.A., Faizy, T.D., et al., 2021. Heterogeneity of multiple sclerosis
Oladosu, O., Liu, W.Q., Pike, B.G., Koch, M., Metz, L.M., Zhang, Y., 2021. Advanced lesions in fast diffusional kurtosis imaging. Bergsland N, ed. PLOS ONE. https://doi.
Analysis of Diffusion Tensor Imaging Along With Machine Learning Provides New org/10.1371/journal.pone.0245844.
Sensitive Measures of Tissue Pathology and Intra-Lesion Activity in Multiple Thompson, A.J., Banwell, B.L., Barkhof, F., et al., 2018. Diagnosis of multiple sclerosis:
Sclerosis. Front. Neurosci. 15, 634063 https://doi.org/10.3389/fnins.2021.634063. 2017 revisions of the McDonald criteria. Lancet Neurol. 17 (2), 162–173. https://
Ontaneda, D., Sati, P., Raza, P., et al., 2021. Central vein sign: A diagnostic biomarker in doi.org/10.1016/S1474-4422(17)30470-2.
multiple sclerosis (CAVS-MS) study protocol for a prospective multicenter trial. Tousignant A, Lemaître P, Precup D, Arnold DL, Arbel T. Prediction of Disease
NeuroImage Clin. 32, 102834 https://doi.org/10.1016/j.nicl.2021.102834. Progression in Multiple Sclerosis Patients using Deep Learning Analysis of MRI Data.
Pinto, M.F., Oliveira, H., Batista, S., et al., 2020. Prediction of disease progression and Accessed November 24, 2021. https://proceedings.mlr.press/v102/tousignant19a.
outcomes in multiple sclerosis with machine learning. Sci. Rep. 10 (1), 21038. html.
https://doi.org/10.1038/s41598-020-78212-6. Tranfa, M., Tortora, M., Pontillo, G., et al., 2022. The central vein sign helps in
Preziosa, P., Pagani, E., Meani, A., et al., 2022. Slowly Expanding Lesions Predict 9-Year differentiating multiple sclerosis from its mimickers: lessons from Fabry disease. Eur.
Multiple Sclerosis Disease Progression. Neurol - Neuroimmunol Neuroinflammation. Radiol. 32 (6), 3846–3854. https://doi.org/10.1007/s00330-021-08487-4.
9 (2), e1139 https://doi.org/10.1212/NXI.0000000000001139. Treaba, C.A., Granberg, T.E., Sormani, M.P., et al., 2019. Longitudinal Characterization
Reich DS, Lucchinetti CF, Calabresi PA. Multiple Sclerosis. Longo DL, ed. N Engl J Med. of Cortical Lesion Development and Evolution in Multiple Sclerosis with 7.0-T MRI.
2018;378(2):169-180. doi:10.1056/NEJMra1401483. Radiology 291 (3), 740–749. https://doi.org/10.1148/radiol.2019181719.
Roca, P., Attye, A., Colas, L., et al., 2020. Artificial intelligence to predict clinical Treaba, C.A., Conti, A., Klawiter, E.C., et al., 2021;3(3):fcab134.. Cortical and phase rim
disability in patients with multiple sclerosis using FLAIR MRI. Diagn Interv Imaging. lesions on 7 T MRI as markers of multiple sclerosis disease progression. Brain
101 (12), 795–802. https://doi.org/10.1016/j.diii.2020.05.009. Commun. https://doi.org/10.1093/braincomms/fcab134.
Samaraweera, A.P.R., Clarke, M.A., Whitehead, A., et al., 2017. The Central Vein Sign in Valcarcel, A.M., Linn, K.A., Khalid, F., et al., 2018. A dual modeling approach to
Multiple Sclerosis Lesions Is Present Irrespective of the T2* Sequence at 3 T: The automatic segmentation of cerebral T2 hyperintensities and T1 black holes in
Central Vein Sign in MS. J. Neuroimaging 27 (1), 114–121. https://doi.org/ multiple sclerosis. NeuroImage Clin. 20, 1211–1221. https://doi.org/10.1016/j.
10.1111/jon.12367. nicl.2018.10.013.
Sati, P., George, I.C., Shea, C.D., Gaitán, M.I., Reich, D.S., 2012. FLAIR*: A Combined MR Valcarcel, A.M., Linn, K.A., Vandekar, S.N., et al., 2018. MIMoSA: An Automated Method
Contrast Technique for Visualizing White Matter Lesions and Parenchymal Veins. for Intermodal Segmentation Analysis of Multiple Sclerosis Brain Lesions: Method
Radiology 265 (3), 926–932. https://doi.org/10.1148/radiol.12120208. For Inter-Modal Segmentation Analysis. J. Neuroimaging 28 (4), 389–398. https://
Sati, P., Thomasson, D., Li, N., et al., 2014. Rapid, high-resolution, whole-brain, doi.org/10.1111/jon.12506.
susceptibility-based MRI of multiple sclerosis. Mult Scler J. 20 (11), 1464–1470. Valverde, J.M., Imani, V., Abdollahzadeh, A., et al., 2021. Transfer Learning in Magnetic
https://doi.org/10.1177/1352458514525868. Resonance Brain Imaging: A Systematic Review. J Imaging. 7 (4), 66. https://doi.
Sati, P., Oh, J., Constable, R.T., et al., 2016. The central vein sign and its clinical org/10.3390/jimaging7040066.
evaluation for the diagnosis of multiple sclerosis: a consensus statement from the Valverde, S., Salem, M., Cabezas, M., et al., 2019. One-shot domain adaptation in
North American Imaging in Multiple Sclerosis Cooperative. Nat Rev Neurol. 12 (12), multiple sclerosis lesion segmentation using convolutional neural networks.
714–722. https://doi.org/10.1038/nrneurol.2016.166. NeuroImage Clin. 21, 101638 https://doi.org/10.1016/j.nicl.2018.101638.
Scalfari, A., Romualdi, C., Nicholas, R.S., et al., 2018. The cortical damage, early Vrenken H, Jenkinson M, Pham D, et al. Opportunities for Understanding MS
relapses, and onset of the progressive phase in multiple sclerosis. Neurology. 90 (24), Mechanisms and Progression With MRI Using Large-Scale Data Sharing and Artificial
e2107–e2118. https://doi.org/10.1212/WNL.0000000000005685. Intelligence. Neurology. Published online October 4, 2021:10.1212/
Simonyan K, Zisserman A. Very Deep Convolutional Networks for Large-Scale Image WNL.0000000000012884. doi:10.1212/WNL.0000000000012884.
Recognition. ArXiv14091556 Cs. Published online April 10, 2015. Accessed May 7, Wattjes, M.P., Ciccarelli, O., Reich, D.S., et al., 2021. 2021 MAGNIMS–CMSC–NAIMS
2020. http://arxiv.org/abs/1409.1556. consensus recommendations on the use of MRI in patients with multiple sclerosis.
Sinnecker, T., Dorr, J., Pfueller, C.F., et al., 2012. Distinct lesion morphology at 7-T MRI Lancet Neurol. 20 (8), 653–670. https://doi.org/10.1016/S1474-4422(21)00095-8.
differentiates neuromyelitis optica from multiple sclerosis. Neurology. 79 (7), Ye, Z., George, A., Wu, A.T., et al., 2020. Deep learning with diffusion basis spectrum
708–714. https://doi.org/10.1212/WNL.0b013e3182648bc8. imaging for classification of multiple sclerosis lesions. Ann. Clin. Transl. Neurol. 7
Sinnecker, T., Clarke, M.A., Meier, D., et al., 2019. Evaluation of the Central Vein Sign as (5), 695–706. https://doi.org/10.1002/acn3.51037.
a Diagnostic Imaging Biomarker in Multiple Sclerosis. JAMA Neurol. 76 (12), 1446. Zeng, C., Gu, L., Liu, Z., Zhao, S., 2020. Review of Deep Learning Approaches for the
https://doi.org/10.1001/jamaneurol.2019.2478. Segmentation of Multiple Sclerosis Lesions on Brain MRI. Front Neuroinformatics.
Solomon, A.J., Schindler, M.K., Howard, D.B., et al., 2016. “Central vessel sign” on 3T 14, 610967 https://doi.org/10.3389/fninf.2020.610967.
FLAIR* for the differentiation of multiple sclerosis from migraine. Ann. Clin. Transl. Zhang H, Zhang J, Yang G, Spincemaille P, Nguyen TD, Wang Y. Memory U-Net:
Neurol. 3 (2), 82–87. https://doi.org/10.1002/acn3.273. Memorizing Where to Vote for Lesion Instance Segmentation. In: ; 2021. Accessed
Solomon, A.J., Bourdette, D.N., Cross, A.H., et al., 2016. The contemporary spectrum of November 19, 2021. https://openreview.net/forum?id=JbWMYLN5Hba.
multiple sclerosis misdiagnosis: A multicenter study. Neurology. 87 (13), Zhang, H., Nguyen, T.D., Zhang, J., et al., 2022. QSMRim-Net: Imbalance-aware learning
1393–1399. https://doi.org/10.1212/WNL.0000000000003152. for identification of chronic active multiple sclerosis lesions on quantitative
Solomon, A.J., Watts, R., Ontaneda, D., Absinta, M., Sati, P., Reich, D.S., 2018. susceptibility maps. NeuroImage Clin. 34, 102979 https://doi.org/10.1016/j.
Diagnostic performance of central vein sign for multiple sclerosis with a simplified nicl.2022.102979.
three-lesion algorithm. Mult Scler J. 24 (6), 750–757. https://doi.org/10.1177/
1352458517726383.

14