COVID Detection From Chest X-Ray Images Using Multi-Scale Attention.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Deep learning shows promise for COVID-19 detection in chest X-rays (CXRs).
• Explainability in these deep learning models for CXR analysis is underexplored.
• Accurate differentiation between COVID-19, non-COVID pneumonia (NCP), and normal cases is crucial.

Purpose of the Study:

• To develop and validate a hierarchical deep learning approach for classifying normal, NCP, and COVID-19 cases using CXR images.
• To incorporate clinical consistency and explainability into the automated detection of COVID-19.
• To evaluate a novel multi-scale attention architecture and a conicity-based loss function.

Main Methods:

• A two-stage hierarchical classification strategy was employed.
• Stage 1: DenseNet-based model to differentiate pneumonia from normal cases.
• Stage 2: Multi-scale Attention Residual Learning (MARL) architecture to distinguish COVID-19 from NCP, trained with a conicity-based loss function.

Main Results:

• The proposed method achieved high classification accuracies across three public datasets: 93% (normal vs. NCP vs. COVID-19), 96.28% (normal vs. NCP), and 84.51% (NCP vs. COVID-19).
• GradCAM attributions provided clinically consistent explanations, aligning with expert annotations.
• MARL highlighted peripheral lung regions for COVID-19 and central regions for NCP, consistent with radiological findings.

Conclusions:

• The developed hierarchical classification approach with the MARL architecture offers competitive performance for COVID-19 detection from CXRs.
• The method provides clinically meaningful explanations, enhancing trust and utility in diagnostic settings.
• The findings underscore the potential of explainable AI in medical image analysis for infectious diseases.