Prototype bank-driven test-time adaptation for medical ultrasound image segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning models excel at medical image segmentation, especially in ultrasound.
• Domain shift, caused by variations in devices and protocols, degrades model performance in real-world clinical settings.
• Existing solutions for domain shift are limited by the need for labeled data or source domain access.

Purpose of the Study:

• To develop a test-time adaptation (TTA) framework for robust ultrasound image segmentation.
• To eliminate the need for source data or target labels in TTA.
• To ensure robustness against distributional drift and prevent catastrophic forgetting.

Main Methods:

• Proposing Prototype Bank-Driven Test-Time Adaptation (PBTTA), a novel TTA framework.
• Utilizing a Dynamic Statistics Fusion Module (DSFM) for domain-level adaptation via fused statistics.
• Employing a Prototype Bank-Guided Semantic Adaptation Module (PBSAM) for semantic-level adaptation using a dynamic prototype bank.
• Implementing a dual-classifier strategy without requiring backpropagation for efficient adaptation.

Main Results:

• PBTTA achieves state-of-the-art performance in ultrasound breast and thyroid tumor segmentation.
• Significant improvements in Dice scores: 15.04% for breast tumors (to 64.82%) and 8.88% for thyroid tumors (to 57.45%).
• Demonstrated excellent robustness against continuous domain shifts and effective mitigation of catastrophic forgetting.

Conclusions:

• PBTTA offers an effective solution for domain shift in ultrasound image segmentation.
• The framework provides robust and accurate segmentation without requiring source data or target labels.
• PBTTA shows significant clinical potential for improving diagnostic accuracy in ultrasound imaging.