Self-Supervised Representation Learning for Ultrasound Video.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Deep learning models for medical image analysis typically require extensive expert annotations.
• Collecting these annotations is costly and challenging, especially for medical imaging data.
• There is a growing need for methods that can learn from unlabelled medical data.

Purpose of the Study:

• To develop a self-supervised learning approach for extracting meaningful representations from unlabelled medical imaging videos.
• To enable learning without human annotations, addressing the scarcity and cost of expert labels.
• To create transferable representations for various medical imaging analysis tasks.

Main Methods:

• Proposing a self-supervised learning framework for medical imaging video.
• Designing the model to perform anatomy-aware tasks using free supervision from the data itself.
• Implementing tasks such as correcting reshuffled video clip order and predicting geometric transformations.

Main Results:

• The proposed self-supervised approach effectively learns meaningful and transferable representations from unlabelled medical imaging video.
• Experimental results on fetal ultrasound videos demonstrate strong performance.
• The learned representations show good transferability to downstream tasks like standard plane detection and saliency prediction.

Conclusions:

• Self-supervised learning is a viable and effective strategy for medical image analysis when labelled data is scarce.
• Anatomy-aware tasks can guide the learning of robust representations from unlabelled medical videos.
• The developed method offers a promising direction for reducing reliance on manual annotations in medical AI.