Machine-Learning-Enabled Diagnostics with Improved Visualization of Disease Lesions in Chest X-ray Images

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Class Activation Maps (CAM) aid in understanding Convolutional Neural Network (CNN) decisions in medical imaging.
• Existing deep learning models for chest X-ray (CXR) analysis often lack explainability through saliency maps.
• Accurate diagnosis of COVID-19 and pneumonia from CXRs is crucial for patient care.

Purpose of the Study:

• To develop an explainable deep learning model for classifying COVID-19, pneumonia, and normal cases from CXRs.
• To enhance classification accuracy and provide visual interpretability of diagnostic findings.
• To introduce a quantitative measure for assessing infection severity using saliency maps.

Main Methods:

• A VGG-16 based deep learning model incorporating image enhancement, region of interest (ROI) cropping, and data augmentation.
• Integration of a Multi-layer Gradient Class Activation Mapping (ML-Grad-CAM) algorithm for generating class-specific saliency maps.
• Definition and calculation of a Severity Assessment Index (SAI) from ML-Grad-CAM outputs.

Main Results:

• The model achieved a 96.44% accuracy for classifying COVID-19, pneumonia, and normal CXRs.
• ML-Grad-CAM generated detailed saliency maps, offering improved visualization compared to standard Grad-CAM.
• The Severity Assessment Index (SAI) provided a quantitative measure of infection severity.

Conclusions:

• The proposed explainable AI approach significantly improves CXR classification accuracy for respiratory diseases.
• ML-Grad-CAM offers valuable visual insights for medical practitioners, aiding in diagnosis and severity assessment.
• This research bridges the gap in explainable AI for medical image analysis, particularly for COVID-19 and pneumonia detection.