Classification and Gradient-based Localization of Chest Radiographs

Contents

• Introduction
• Overview
• Steps
• Results
• Usage
• Conclusions
• References

Introduction

A team of radiologists from New Orleans studied the usefulness of Chest Radiographs for diagnosing COVID-19 compared to the reverse-transcription polymerase chain reaction (RT-PCR) and found out they could aid rapid diagnosis, especially in areas with limited testing facilities [1].
Another study found out that the radiographs of different viral cases of pneumonia are comparative, and they overlap with other infectious and inflammatory lung diseases, making it hard for radiologists to recognize COVID‐19 from other viral pneumonia cases [2].
This project aims to make the former study a reality while dealing with the intricacies in the latter, with the help of Deep Learning.

Overview

The project uses the COVID-19 Radiography Database [3] as it's dataset. It has a total of 21165 Chest X-Rays (CXRs) belonging to 4 different classes (COVID-19, Lung Opacity, Normal and Viral Pneumonia).
Three top scoring CNN architectures, VGG-16 [4], ResNet-18 [5] and DenseNet-121 [6], trained on the ImageNet Dataset [7], were chosen for fine-tuning on the dataset.
The results obtained from the different architectures were then evaluted and compared.
Finally, with the help of Gradient weighted Class Activation Maps (Grad-CAM) [8] the affected areas in CXRs were localized.

• Note: The dataset and the trained models can be found in here.
  

Steps

1. Dataset Exploration
2. Split the dataset

   Type	COVID-19	Lung Opacity	Normal	Viral Pneumonia	Total
   Train	3496	5892	10072	1225	20685
   Val	60	60	60	60	240
   Test	60	60	60	60	240

3. Fine-tune VGG-16, ResNet-18 and DenseNet-121
   1. Define Transformations
   2. Handle imbalanced dataset with Weighted Random Sampling (Over-sampling)
   3. Prepare the Pre-trained models
   4. Fine-tune step with Early-stopping

      • Hyper-parameters
        Learning rate	0.00003
        Batch Size	32
        Number of Epochs	25

      • Loss Function	Optimizer
        Categorical Cross Entropy	Adam

   5. Plot running losses & accuracies

      • Model	Summary Plot
        VGG-16
        ResNet-18
        DenseNet-121

4. Results Evaluation
   1. Plot confusion matrices
   2. Compute test-set Accuracy, Precision, Recall & F1-score
   3. Localize using Grad-CAM
5. Inference

Results

	VGG-16	ResNet-18	DenseNet-121
Pathology COVID-19 Lung Opacity Normal Viral Pneumonia	Accuracy Precision Recall F1-Score 0.9956 0.9833 1.0000 0.9916 0.9582 0.8833 0.9464 0.9138 0.9622 0.9667 0.8923 0.9280 0.9913 0.9833 0.9833 0.9833	Accuracy Precision Recall F1-Score 0.9871 0.9667 0.9830 0.9748 0.9664 0.8667 1.0000 0.9286 0.9664 1.0000 0.8823 0.9375 0.9957 1.0000 0.9836 0.9917	Accuracy Precision Recall F1-Score 0.9957 0.9833 1.0000 0.9916 0.9623 0.9167 0.9322 0.9244 0.9623 0.9500 0.9047 0.9268 0.9957 0.9833 1.0000 0.9916
TL;DR Train set Test set	Total Correct Predictions Total Accuracy 20362 98.44% 229 95.42%	Total Correct Predictions Total Accuracy 20639 99.78% 230 95.83%	Total Correct Predictions Total Accuracy 20540 99.30% 230 95.83%
Confusion Matrices

• Localization with Gradient-based Class Activation Maps

COVID-19 infected CXR	VGG-16	ResNet-18	DenseNet-121

Usage

• Clone the repository

git clone 'https://github.com/priyavrat-misra/xrays-and-gradcam.git' && cd xrays-and-gradcam/

• Install dependencies

pip install -r requirements.txt

• Using argparse script for inference

python overlay_cam.py --help

usage: GradCAM on Chest X-Rays [-h] [-i IMAGE_PATH]
                               [-l {covid_19,lung_opacity,normal,pneumonia}]
                               -m {vgg16,resnet18,densenet121}
                               [-o OUTPUT_PATH]

Overlays given label's CAM on a given Chest X-Ray.

optional arguments:
  -h, --help            show this help message and exit
  -i IMAGE_PATH, --image-path IMAGE_PATH
                        Path to chest X-Ray image.
  -l {covid_19,lung_opacity,normal,pneumonia}, --label {covid_19,lung_opacity,normal,pneumonia}
                        Choose from covid_19, lung_opacity, normal &
                        pneumonia, to get the corresponding CAM. If not
                        mentioned, the highest scoring label is considered.
  -m {vgg16,resnet18,densenet121}, --model {vgg16,resnet18,densenet121}
                        Choose from vgg16, resnet18 or densenet121.
  -o OUTPUT_PATH, --output-path OUTPUT_PATH
                        Format: "<path> + <file_name> + .jpg"

• An example

python overlay_cam.py --image-path ./assets/original.jpg --label covid_19 --model resnet18 --output-path ./assets/dense_cam.jpg

GradCAM generated for label "covid_19".
GradCAM masked image saved to "./assets/res_cam.jpg".

Conclusions

• DenseNet-121 having only 7.98 Million parameters did relatively better than VGG-16 and ResNet-18, with 138 Million and 11.17 Million parameters respectively.
• Increase in model's parameter count doesn’t necessarily achieve better results, but increase in residual connections might.
• Oversampling helped in dealing with imbalanced data to a great extent.
• Fine-tuning helped substantially by dealing with the comparatively small dataset and speeding up the training process.
• GradCAM aided in localizing the areas in CXRs that decides a model's predictions.
• The models did a good job distinguishing various infectious and inflammatory lung diseases, which is rather hard manually, as mentioned earlier.

References

• [1] David L. Smith, John-Paul Grenier, Catherine Batte, and Bradley Spieler. A Characteristic Chest Radiographic Pattern in the Setting of the COVID-19 Pandemic. Radiology: Cardiothoracic Imaging 2020 2:5.
• [2] Hyun Jung Koo, Soyeoun Lim, Jooae Choe, Sang-Ho Choi, Heungsup Sung, and Kyung-Hyun Do. Radiographic and CT Features of Viral Pneumonia. RadioGraphics 2018 38:3, 719-739.
• [3] Tawsifur Rahman, Muhammad Chowdhury, Amith Khandakar. COVID-19 Radiography Database. Kaggle.
• [4] Karen Simonyan, Andrew Zisserman. Very Deep Convolutional Networks for Large-Scale Image Recognition. arxiv:1409.1556v6.
• [5] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun. Deep Residual Learning for Image Recognition. arxiv:1512.03385v1.
• [6] Gao Huang, Zhuang Liu, Laurens van der Maaten, Kilian Q. Weinberger. Densely Connected Convolutional Networks. arxiv:1608.06993v5.
• [7] Deng, J. et al., 2009. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition. pp. 248–255.
• [8] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, Dhruv Batra. Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization. arXiv:1610.02391v4.