Weakly-supervised lesion analysis with a CNN-based framework for COVID-19.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Pulmonology

Background:

• Chest CT scans are crucial for diagnosing COVID-19, but lesion analysis is complex.
• Manual delineation of diverse COVID-19 lesions is challenging and time-consuming.
• Existing methods often require detailed annotations, limiting their applicability.

Purpose of the Study:

• To develop a weakly-supervised framework for automated COVID-19 diagnosis and lesion analysis using chest CT images.
• To enable lesion identification without requiring specific location or type annotations.
• To integrate diagnosis and lesion feature extraction into a unified model.

Main Methods:

• A deep learning framework with separate diagnosis and lesion identification branches was employed.
• The model was trained using a weakly-supervised approach with only normal and abnormal CT images.
• The framework was validated on public datasets and clinical data from 13 patients.

Main Results:

• The proposed framework achieved state-of-the-art performance in COVID-19 diagnosis prediction.
• Extracted lesion features effectively distinguished between ground glass opacity and consolidation.
• The model demonstrated strong diagnostic accuracy and lesion characterization capabilities.

Conclusions:

• The weakly-supervised framework offers an effective method for COVID-19 diagnosis and lesion analysis from CT scans.
• The approach eliminates the need for pixel-wise annotations, simplifying the analysis process.
• The framework shows potential for discovering novel lesion types and generalizing to other chest diseases.