Context Matters: Graph-based Self-supervised Representation Learning for Medical Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Science

Background:

• Supervised learning for medical imaging requires extensive annotated datasets, which are costly and time-consuming to acquire.
• Existing self-supervised methods often fail to adequately incorporate the anatomical context crucial for medical image analysis.
• Limited annotated COVID-19 imaging datasets hinder the development of effective diagnostic tools.

Purpose of the Study:

• To develop a novel self-supervised learning framework that leverages anatomical context for medical image analysis.
• To improve the representation learning for medical images, particularly in the context of COVID-19.
• To create a method capable of quantifying disease progression and generalizing across different patient cohorts.

Main Methods:

• Introduced a two-level self-supervised representation learning approach incorporating regional anatomical and patient-level objectives.
• Utilized graph neural networks to model relationships between anatomical regions, informed by an anatomical atlas.
• Handled arbitrarily sized medical images at full resolution using graph representation.

Main Results:

• The proposed method demonstrated superior performance compared to baseline approaches lacking contextual information on large-scale lung CT datasets.
• The learned embeddings effectively quantified the clinical progression of COVID-19.
• The model exhibited strong generalization capabilities across COVID-19 patients from diverse hospital settings.

Conclusions:

• The novel self-supervised learning approach effectively incorporates anatomical context, outperforming traditional methods.
• The method provides a robust tool for quantifying COVID-19 progression and shows promise for clinical applications.
• The model's ability to identify clinically relevant regions suggests its potential for aiding in medical image interpretation.