Exploiting Shared Knowledge From Non-COVID Lesions for Annotation-Efficient COVID-19 CT Lung Infection Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• The COVID-19 pandemic necessitates accurate quantitative analysis of lung infections using computed tomography (CT).
• Automatic lung infection segmentation in CT scans is crucial but hindered by a scarcity of annotated COVID-19 datasets.
• Existing public non-COVID lung lesion datasets offer potential for knowledge transfer.

Purpose of the Study:

• To develop an annotation-efficient model for COVID-19 CT lung infection segmentation.
• To exploit shared knowledge from non-COVID lung lesions to enhance segmentation performance.
• To improve the generalizability and discriminative power of deep learning models for COVID-19 analysis.

Main Methods:

• A novel relation-driven collaborative learning model with a general encoder for non-COVID lesions and a target encoder for COVID-19 infections.
• A collaborative learning scheme to enforce feature-level relation consistency.
• Training on limited COVID-19 data augmented with knowledge from diverse non-COVID lesion datasets.

Main Results:

• Improved state-of-the-art performance in COVID-19 segmentation, with up to 3.0% increase in Dice Similarity Coefficient and 4.2% in Normalized Surface Dice.
• Significant outperformance of cutting-edge segmentation methods on large-scale 2D CT datasets.
• Demonstrated effectiveness of leveraging non-COVID data for annotation-efficient learning.

Conclusions:

• Exploiting shared knowledge from non-COVID lesions significantly enhances COVID-19 CT segmentation accuracy with limited annotations.
• The proposed model offers a promising approach for annotation-efficient deep learning in medical image analysis.
• The method shows strong potential for real-world applications in global COVID-19 response efforts.