Generating and Weighting Semantically Consistent Sample Pairs for Ultrasound Contrastive Learning.

Area of Science:

• Medical Imaging
• Machine Learning
• Computer-Aided Diagnosis

Background:

• Deep neural networks (DNNs) require well-annotated medical datasets for effective lesion feature extraction, but creating these is costly and expertise-intensive.
• Pre-training DNNs on ImageNet is common for limited data but introduces a domain gap between natural and medical images.
• Ultrasound (US) imaging presents unique challenges due to domain-specific characteristics.

Purpose of the Study:

• To develop a self-supervised learning method for pre-training DNNs on unlabeled ultrasound videos, reducing the domain gap in medical applications.
• To improve the generalization capabilities of DNNs for various computer-aided diagnosis (CAD) tasks using medical ultrasound data.
• To address the challenge of creating semantically consistent sample pairs for contrastive learning in medical imaging.

Main Methods:

• Proposed a novel meta-learning-based contrastive learning method, Meta Ultrasound Contrastive Learning (Meta-USCL), for learning US image representations from unlabeled US videos.
• Introduced a positive pair generation module to ensure semantically consistent samples for contrastive learning.
• Implemented an automatic sample weighting module leveraging meta-learning principles.

Main Results:

• Meta-USCL achieved state-of-the-art (SOTA) performance across multiple computer-aided diagnosis (CAD) tasks.
• Demonstrated effectiveness in pneumonia detection, breast cancer classification, and breast tumor segmentation.
• Successfully reduced the domain gap by pre-training on ultrasound domains instead of ImageNet.

Conclusions:

• The proposed self-supervised Meta-USCL method is highly effective for medical ultrasound image analysis.
• Meta-USCL offers a viable solution for leveraging unlabeled ultrasound data, overcoming limitations of traditional pre-training methods.
• The approach shows significant potential for advancing computer-aided diagnosis in various medical fields.