Adaptive batch-fusion self-supervised learning for ultrasound image pretraining.

Area of Science:

• Medical Imaging
• Machine Learning
• Computer Vision

Background:

• Medical self-supervised learning (SSL) offers efficient feature extraction without manual labels.
• Current SSL methods face challenges due to complex pretext tasks and over-reliance on data augmentation.
• This dependency creates a bottleneck in advancing medical SSL research.

Purpose of the Study:

• To re-evaluate the role of data augmentation in medical image SSL feature learnability.
• To introduce an adaptive data augmentation strategy using batch fusion.
• To propose a convolutional embedding block for incremental batch representation learning.

Main Methods:

• Developed an adaptive self-supervised learning data augmentation method based on batch fusion.
• Introduced a convolutional embedding block to learn incremental representations between batches.
• Tested the method on five fused data tasks using Vision Transformer (ViT) with a linear classification protocol.

Main Results:

• Achieved a 94.25% linear classification accuracy with only 150 self-supervised feature training steps in ViT.
• Outperformed existing methods in the same category.
• Ablation studies confirmed the effectiveness of the proposed augmentation strategy for ultrasound data features in SSL.

Conclusions:

• The proposed adaptive data augmentation strategy significantly improves feature representation in medical SSL.
• This method addresses the bottleneck caused by traditional data augmentation dependencies.
• The approach demonstrates high efficiency and performance, particularly for ultrasound image analysis.