Multiscale Switch for Semi-Supervised and Contrastive Learning in Medical Ultrasound Image Segmentation

Area of Science:

• Medical imaging
• Computer vision
• Machine learning

Background:

• Medical ultrasound (US) image segmentation is challenging due to limited labeled data and artifacts like speckle noise.
• Existing semi-supervised learning (SSL) methods struggle with effective unlabeled data use and robust feature extraction.

Purpose of the Study:

• To introduce Switch, a novel SSL framework designed to overcome limitations in US image segmentation.
• To enhance feature representation and unlabeled data utilization for improved segmentation accuracy.

Main Methods:

• Switch employs a multiscale switch (MSS) strategy for uniform spatial coverage via hierarchical patch mixing.
• A frequency-domain switch (FDS) with contrastive learning performs amplitude switching in Fourier space for robust features.
• The framework integrates these innovations within a teacher-student architecture.

Main Results:

• Switch demonstrates superior performance across six diverse US datasets (lymph nodes, breast lesions, thyroid nodules, prostate).
• At a 5% labeling ratio, Switch achieved significant Dice score improvements (e.g., 80.04% on LN-INT, 85.52% on DDTI).
• The semi-supervised approach surpassed fully supervised baselines, showing remarkable efficiency with 1.8M parameters.

Conclusions:

• Switch offers a parameter-efficient and effective solution for medical US image segmentation, particularly in low-data scenarios.
• The proposed MSS and FDS strategies significantly improve segmentation accuracy and feature robustness.
• This framework is highly suitable for resource-constrained medical imaging applications.