Project Name: A Hybrid Multi-stage Network for Lung Segmentation, Disease Classification and Severity Localization from X-ray Images

GitHub Folder Structure

• segmentation-notebook.ipynb - Contains the code and implementation of lung image segmentation.
• classification-and-localisation-notebook.ipynb - Contains the code and implementation of lung disease classification and localalisation.

How to run the notebook

Prerequisites

Before starting, ensure you have a Kaggle account or access to Google Colab, and Python 3 installed if running locally.

• Step 1: Download the Dataset

  • Download the COVID-19 Radiography Database.
  • The dataset should include images and masks for COVID-19, Lung Opacity, Normal, and Viral Pneumonia.

• Step 2: Setup Your Environment

  • Kaggle: Upload the dataset to your Kaggle account and use it in a new notebook.
  • Google Colab: Upload the dataset to Google Drive, mount the drive in Colab, and update paths accordingly:

• Step 3: Update the paths to the dataset in your notebook based on your environment setup.

• Step 4: Run all the cells

DATASET

For our project, we are using the COVID-19 Radiography Database. This comprehensive database contains chest X-ray images for three distinct classes: COVID-19, normal, and viral pneumonia. Specifically, the dataset comprises 3616 images of COVID-19 positive cases, 10,192 images categorized as normal, and 1345 images identified as viral pneumonia. This extensive collection allows us to train our diagnostic models effectively, ensuring robust performance in identifying and classifying these conditions.

ABOUT

Accurate and rapid diagnosis of respiratory diseases such as COVID-19 and viral pneumonia using chest X-rays (CXRs) is crucial for timely treatment and containment efforts. However, traditional diagnostic approaches often struggle with high variability in image quality and the subtlety of disease manifestations, leading to a significant rate of diagnostic errors. To address these challenges, this report presents a novel hybrid multi-stage network that initially segments the lung region in the CXR images, followed by classification and subsequent localization of the disease using Grad-CAM. This approach allows for focused analysis on relevant lung areas, enhancing the model's accuracy and reliability in diagnosing respiratory conditions

METHODOLOGY

Our approach involves a systematic progression through three stages: segmentation, classification, and localization. This structured workflow allows us to precisely isolate and analyze lung regions, identify pathological conditions, and visually highlight critical areas influencing diagnostic outcomes, thereby facilitating a comprehensive examination of CXR images.

RESULTS

Segmentation Models Experiment Results

UNet

Backbone	binary_accuracy	dice_coef	iou_score
Mobilenetv2	0.7585	0.5252	0.7511
seresnext50	0.9886	0.9757	0.9941
inceptionresnetv2	0.9906	0.9794	0.995
efficientnetb2	0.9894	0.9769	0.9945
vgg16	0.9282	0.7714	0.9403

LinkNet

Backbone	binary_accuracy	dice_coef	iou_score
Mobilenetv2	0.6167	0.4094	0.7271
seresnext50	0.9967	0.9925	0.9981
inceptionresnetv2	0.998	0.9946	0.9985
efficientnetb2	0.9905	0.9794	0.995
vgg16	0.989	0.969	0.9907

FPN

Backbone	binary_accuracy	dice_coef	iou_score
Mobilenetv2	0.6587	0.4794	0.7374
seresnext50	0.998	0.9946	0.9985
inceptionresnetv2	0.9913	0.9818	0.9955
efficientnetb2	0.9920	0.9826	0.9957
vgg16	0.9911	0.972	0.9917

Classification Models Experiment Results and graphs:

Model	Val Loss	Val Acc.	Test Loss	Test Acc.
CoAtNet0	0.159	95.8%	0.144	95.49%
Xception	0.182	95.63%	0.220	93.80%
ResNet50	0.182	93.11%	0.220	92.64%
InceptionResNet50	0.193	94.81%	0.197	93.87%

Grad-CAM Visualizations