COVID-19 disease identification network based on weakly supervised feature selection.

Area of Science:

• Medical Imaging and Artificial Intelligence
• Computer-Aided Diagnosis
• Deep Learning in Healthcare

Background:

• The COVID-19 pandemic presents significant global health challenges, straining healthcare systems.
• Artificial intelligence (AI) offers potential for rapid and accurate diagnosis, especially during outbreaks with limited resources.
• Computed Tomography (CT) imaging is crucial for COVID-19 detection, but interpretation requires expertise.

Purpose of the Study:

• To propose a weakly supervised deep learning network, W-COVNet, for accurate COVID-19 classification from CT images.
• To enhance diagnostic efficiency and accuracy in the context of the COVID-19 pandemic.
• To develop a model capable of feature selection, fusion, and visualization for improved interpretability.

Main Methods:

• Developed a three-module network: weakly supervised feature selection (W-FS), deep learning bilinear feature fusion (DBFF), and Grad-CAM++ based visualization (Grad-V).
• W-FS module focused on selecting relevant features from CT images, removing background noise.
• DBFF module employed symmetric networks for feature extraction and fusion, while Grad-V enabled lesion visualization in unlabeled images.

Main Results:

• The W-COVNet achieved an average classification accuracy of 85.3% via fivefold cross-validation.
• Comparative analysis demonstrated superior performance of W-COVNet against seven advanced classification models.
• The Grad-V module successfully visualized lesions, aiding in the interpretation of results.

Conclusions:

• The proposed W-COVNet is an effective tool for COVID-19 classification from CT images.
• Weakly supervised learning and feature fusion techniques contribute to improved diagnostic accuracy.
• W-COVNet shows promise for assisting healthcare professionals in managing the COVID-19 pandemic.