Benchmarking Self-Supervised Representation Learning from a million Cardiac Ultrasound images.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Supervised deep learning is standard but requires extensive annotated medical data.
• Self-supervised learning (SSL) uses unlabeled data, reducing annotation overhead.
• SSL methods achieve near-supervised performance on natural images.

Purpose of the Study:

• Compare state-of-the-art SSL methods for medical image analysis.
• Evaluate SSL performance on challenging cardiac view classification from ultrasound.
• Analyze the impact of data size on SSL training phases.

Main Methods:

• Benchmarked various state-of-the-art self-supervised learning techniques.
• Assessed performance on cardiac ultrasound view classification task.
• Investigated the influence of dataset size during pre-text and main task training.
• Compared SSL approaches with a task-specific SSL method and ImageNet transfer learning.

Main Results:

• SSL methods demonstrate competitive performance in medical imaging tasks.
• Performance is influenced by the amount of data used in both pre-text and main training phases.
• SSL offers a viable alternative to supervised learning, especially when labeled data is scarce.

Conclusions:

• Self-supervised learning is effective for cardiac ultrasound view classification.
• Optimizing data size is crucial for maximizing SSL performance in medical applications.
• SSL presents a scalable solution for leveraging large unlabeled medical imaging datasets.