A novel collaborative self-supervised learning method for radiomic data.

Area of Science:

• Medical Imaging
• Machine Learning
• Radiomics

Background:

• Computer-aided disease diagnosis from radiomic data is crucial but hindered by the costly process of labeling radiological images.
• Existing methods struggle with the unique characteristics of radiomic data compared to text or standard images.

Purpose of the Study:

• To present a novel collaborative self-supervised learning method to address the challenge of insufficient labeled radiomic data.
• To reduce human annotation efforts in developing disease diagnosis techniques.

Main Methods:

• Developed a collaborative self-supervised learning framework tailored for radiomic data.
• Introduced two pretext tasks to explore latent pathological relationships and inter-subject data similarities/dissimilarities.
• Learned robust latent feature representations from unlabeled radiomic data.

Main Results:

• The proposed method outperformed state-of-the-art self-supervised learning techniques on both classification and regression tasks.
• Demonstrated superior performance on a simulation study and two independent datasets.
• Validated the effectiveness in reducing reliance on human annotation for disease diagnosis.

Conclusions:

• The novel collaborative self-supervised learning method effectively learns feature representations from radiomic data.
• This approach shows significant potential for automatic disease diagnosis using large-scale unlabeled datasets.
• Further refinement could enhance its utility in clinical settings.