Shifting to machine supervision: annotation-efficient semi and self-supervised learning for automatic medical image...

Area of Science:

• Medical Imaging Analysis
• Machine Learning in Healthcare
• Artificial Intelligence in Medicine

Background:

• Supervised learning for clinical treatment advancements is limited by the need for extensive annotated medical data.
• Manual data annotation is time-consuming and expensive, requiring significant clinical specialist involvement.

Purpose of the Study:

• To introduce and evaluate the S4MI (Self-Supervision and Semi-Supervision for Medical Imaging) pipeline.
• To assess the effectiveness of self-supervised and semi-supervised learning in medical image classification and segmentation tasks.
• To demonstrate a more scalable and efficient alternative to fully-supervised methods in medical AI.

Main Methods:

• The study utilized the S4MI pipeline, integrating self-supervised and semi-supervised learning techniques.
• These methods were benchmarked on three diverse medical imaging datasets.
• Performance was evaluated for both image classification and segmentation tasks.

Main Results:

• Self-supervised learning outperformed fully-supervised methods in all evaluated medical image classification tasks.
• Semi-supervised learning achieved superior segmentation results compared to fully-supervised methods, using 50% fewer labels.
• The S4MI pipeline demonstrated the potential to reduce annotation burden while enhancing model performance.

Conclusions:

• Self-supervised and semi-supervised learning offer powerful alternatives to traditional supervised methods in medical imaging.
• The S4MI pipeline provides a scalable solution for machine supervision, reducing the need for extensive manual annotation.
• Open-sourcing the S4MI code facilitates broader adoption and further research in AI-driven medical diagnostics.