Detecting and Recognising

Lung Cancer
Using Convolutional Neural Networks

Major Project by

161112001 Abhishek Pandey

161112031 Lokesh Lovewanshi
161112046 Sudhanshu Ranjan
161112049 Shubham Kose
AGENDA
Abstract
Introduction
Literature Review
Methodology and Work Description
Tools and Technology to be Used
Implementation and Coding
Result Analysis
Conclusion and Future Scope
Abstract
● Lung cancer is one of the most dreadful diseases in the developing countries and its
mortality rate is 19.4%. Early detection of lung tumor is done by using many imaging
techniques such as Computed Tomography (CT), Sputum Cytology, Chest X-ray and
Magnetic Resonance Imaging (MRI).

● The chance of survival at the advanced stage is less when compared to the treatment
and lifestyle to survive cancer therapy when diagnosed at the early stage of the cancer.
Manual analysis and diagnosis system can be greatly improved with the implementation
of image processing techniques.

● Neural network plays a key role in the recognition of the cancer cells among the normal
tissues, which in turn provides an effective tool for building an assistive AI based cancer
detection. The cancer treatment will be effective only when the tumor cells are
accurately separated from the normal cells.

● Classification of the tumor cells and training of the neural network forms the basis for
the machine learning based cancer diagnosis. This major project presents a
Convolutional Neural Network (CNN) based technique to classify the lung tumors as
malignant or benign.
 Introduction

01 A neural network in a modern sense is a network or circuit of artificial neurons,

to build an artificial neural network. The connections of the neurons are
Neural
Networks modeled as weights. A positive weight reflects an excitatory connection, while
negative values mean inhibitory connections.

02 CNNs, like neural networks, are made up of neurons with learnable weights
and biases. Each neuron receives several inputs, takes a weighted sum over
Convolutional
Neural Networks them, passes it through an activation function and responds with an output.
The whole network has a loss function and all the tips and tricks that was
developed for neural networks still apply on CNNs.

03 ●
●
Image Recognition
Video Analysis
Uses of
CNNs ● Natural Language Processing
● Drug Discovery
 Literature Review
● Computer-Aided Diagnostic (CAD) approaches use a filter for enhancement of lesions as a
preprocessing step for enhancing sensitivity and specificity. Thus, existing filters fail to improve actual
lesions. Suzuki et al (2005) proposed a supervised filter for enhancement of lesions by use of a
Massive-Training Artificial Neural Network (MTANN) in a CAD scheme for detection of lung nodules
in CT. The MTANN filter was trained with actual nodules in CT images to improve actual patterns of
nodules. By use of the MTANN filter, the sensitivity and specificity of this CAD approach were
enhanced. With the database of 69 lung cancers, this CAD approach with the MTANN filter achieved
97% sensitivity with 6.7 false positives (FPs) per section, whereas a conventional CAD technique with
a difference-image technique achieved 96% sensitivity.

● In (Nikita, 2012), a sober edge detection method was used which is based on finding the image
gradient. This method tells that intensity of the image will be maximum where there is a separation of
two dissimilar regions and thus an edge must exist there. On this basis they found the nodules in CT
images. In (Parsh, 2011), a new variation level set algorithm without re-initialization was used. They
also used thresholding to reduce the noise component of the images.

● In (Sonith, 2012) an overview of entire process for processing digital images for lung cancer detection
is given in this paper. This paper also describes all the essential steps required for the better
performance starting from the pre-processing till the very end phase extraction of features
 Literature Review
● Regarding lung cancer diagnosis, methods proposed so far have dealt mostly with radiology. In image-
based radiomics features strongly related to survival are extracted from positron emission
tomography-computed tomography (PET/CT) scans. CNN is employed for classification of lung
nodule images yielding accuracy of 86.4%. In digital pathology tasks CNNs chave been used on cell
level for mitosis detection and cell nuclei detection. CAMELYON16 was the first challenge dealing
with WSIs to detect breast cancer metastases in lymph nodes.

● Thanks to the availability of large annotated training set in this challenge, it was possible to train
deeper and more powerful CNN architectures like GoogLeNet, VGG-Net and ResNet. Method that
gives the best result in this challenge is described in. It performs patch-based classification to
discriminate tumor patches from normal patches using a combination of 2 GoogLeNet architectures
where one of them is trained with and another without hard-negative mining.

● Aim of TUPAC challenge was WSI based mitosis detection in breast cancer tissue and tumor grading
prediction. In the best performing method ROI regions are firstly extracted from WSI based on cell
density. This is followed by mitosis detection using ResNet CNN architecture. Finally, each WSI is
represented by feature vector including the number of mitoses and cells in each patch as well as other
features derived from statistics.
Methodology and Work Description
For the purpose of the project, we are using Kaggle dataset and LUNA dataset .

The CNN will be developed with variable depths to evaluate the performance of these models
for facial expression recognition.

The first part of the network refers to M convolutional layers that can possess spatial batch
normalization, dropout, and max-pooling in addition to the convolution layer and ReLU
nonlinearity, which always exists in these layers. After M convolution layers, the network is led
to N fully connected layers that always have Affine operation and ReLU nonlinearity, and can
include batch normalization and dropout. Finally, the network is followed by the affine layer
that computes the scores and softmax loss function.
Methodology and Work Description

Fig. 1 - Non Cancerous Lung Fig. 2 - Cancerous Lung with Nodule

Methodology and Work Description
The developed model gives the user the freedom to decide about the number of convolutional
and fully connected layers, as well as the existence of batch normalization and max -pooling
layers. Along with batch normalization techniques, regularization was included in the
implementation. Furthermore, the number of filters and layers can be specified by user for the
best results .

Based on the results of previous publications, a decision was taken to create a CNN by oneself
and train it from scratch. A 9-layer CNN with two convolutional layers, two pooling layers, and
4 fully connected dense layers along with a matrix flattening layer. The structure of the CNN is
shown in Fig 3.
Fig. 3 - CNN Architecture Flowchart
Methodology and Work Description
1. Extracting Effective Features
In this module, first the system will take the image from the dataset taken. Then the
input image is first checked for the lung x-ray features. In case if the image does not
contain lung features, then it is not detected. If the input image contains lung features,
then it detects the features. Lung is detected from the image as shown as Fig. 3

Fig. 4 - Image Fed to Model Fig. 5 - Extracting Lung

Methodology and Work Description
2. Feature Point Detection
For lung detection, first - convert the image from an RGB format to a binary format. The next
step is to find the ribs from the binary image. System will start scanning from the middle of the
image, after that it will look for continuous white pixels after a continuous black pixel. In this
the goal is to want to find the maximum width of the white pixel by searching vertical both left
and right side. Then, if the new width is smaller than half of the previous maximum width, then
the scan is broken because in that case the scan will reach the diaphragm. Then the lung is cut
from the starting position of the x ray and its height will be 1.5 multiple of its width. This
processed image will have the lung, hotspot and body.

Fig. 6 - Feature Point Detection

Tools and Technology to be used
1. Software Requirements
a. Python
b. Keras Library
c. Anaconda Navigator
d. Numpy Library

2. Hardware Requirements
a. Windows XP or Above
b. 2GB of RAM
c. Any Dual Core Processor or above
 Implementation and Coding
S. No Layer Shape

1 Convolution2D (Filters(64, 3, 3), input_shape = (64, 64, 3), activation =

‘relu’)
2 MaxPooling2D (pool_size = (2, 2))

3 Convolution2D (Filters(32, 3, 3), activation = ‘relu’)

4 MaxPooling2D (pool_size = (2, 2))

5 Flatten Flatten the matrix

6 Dense (output_dim = 128, activation = ‘relu’)

7 Dense (output_dim = 128, activation = ‘relu’)

8 Dense (output_dim = 128, activation = ‘relu’)

9 Dense (output_dim = 2, activation = ‘softmax’)

Result Analysis
S. No Image Output Accuracy

Correct

Correct
Result Analysis
S. No Image Output Accuracy

Incorrect

Incorrect
Result Analysis
S. No Image Output Accuracy

Correct

Incorrect
Result Analysis
S. No Image Output Accuracy

Correct

Correct
Result Analysis
S. No Image Output Accuracy

Correct

10

Correct
Result Analysis
S. No Image Output Accuracy

11

Correct

12

Incorrect
Result Analysis
S. No Image Output Accuracy

13

Correct

14

Correct
 Result Analysis
S. No Image Output Accuracy

15

Correct

Accuracy on the above images = (Number of correct instances/Number of total instances) x 100

Accuracy on the above images = (11/15) x 100 = 73.33%

Accuracy on training dataset = 98%

Accuracy on the test dataset = 76%

Conclusion and Future Scope
The first novelty in our paper is using the K-means algorithm to pre-classify the pictures into
piles of same slice images, where the DNN can specialize in image classification of same slice
images.

The second novelty is that the additional convolution layer with edge sharpening filters, to
thoroughly search for cancer. Finally, the most novelty is testing our Deep Neural Network
with carcinoma images from Tx stages 2, 3 and 4 and determining at which Tx stage the 2
algorithms can detect the possibility of cancer. The results were analyzed with medical
personnel from the oncology department and were marked as satisfactory to see cancer in T3
phase.

For future work, we plan on making an extra analysis, where we are going to change the DNN
to output 2 values (0 and 1) and determine which one has higher certainty of classification.
This way, we can classify the image not even as being a decimal value between 0.0 or 1.0, but
also compare what proportion is 0 (not cancer) and the way much is 1 (cancer). For extra future
work, almost like Cruz-Roa and Ovalle, who used RGB (color) images to spotlight the realm of
malignant cells, we plan on modifying the DNN to indicate to us where (the location) on the CT
image it's detected a cancer.
Thank you!