A Practical Study of Data Requirements for Self-Supervised Learning in Medical Image Analysis.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning dominates medical image analysis, but annotated datasets are scarce.
• Self-supervised learning (SSL) offers a solution for limited data scenarios.
• Optimal pretraining and fine-tuning strategies for SSL in medical imaging require further investigation.

Purpose of the Study:

• To evaluate the performance of contrastive self-supervised learning (SSL) models in limited data settings.
• To assess the impact of pretraining and fine-tuning dataset sizes on SSL model accuracy for binary classification tasks.
• To identify effective SSL strategies for medical image analysis with scarce data.

Main Methods:

• Evaluated widely adopted contrastive SSL models.
• Assessed model performance based on varying numbers of images for pretraining and fine-tuning.
• Focused on binary classification tasks in medical image analysis.

Main Results:

• Simple Siamese Representation Learning (SimSiam) achieved high accuracy with as few as four training images.
• Effective performance with SimSiam likely necessitates a minimum of 10,000 pretraining images.
• Findings highlight the influence of dataset scale on SSL model efficacy.

Conclusions:

• SSL, particularly SimSiam, can be effective in limited data medical image analysis.
• Substantial pretraining datasets are crucial for achieving high accuracy in SSL fine-tuning.
• These insights aid in optimizing SSL pipelines for rare diseases and data-scarce clinical environments.