SicTTA: Single image continual test time adaptation for medical image segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Test Time Adaptation (TTA) is crucial for model robustness in evolving clinical data.
• Existing TTA methods struggle with limited data and distribution shifts common in medical imaging.
• Pseudo-labeling in TTA can be unreliable due to domain misalignment.

Purpose of the Study:

• To introduce SicTTA, a novel single-image continual test time adaptation method for medical image segmentation.
• To address limitations of current TTA methods in clinical settings, particularly with limited data and distribution shifts.
• To improve the robustness and reliability of segmentation models in unseen target domains.

Main Methods:

• SicTTA utilizes Class Compact Density (CCD) analysis for uncertainty estimation and Source-Friendly Target (SFT) image selection.
• A first-in-first-out strategy manages the SFT image and feature pool for size constraints.
• Source-Aligned Batch Enhancement (SABE) and Similarity-driven Feature Fusion (SFF) adapt test images to the source model's distribution.

Main Results:

• SicTTA significantly outperforms seven state-of-the-art TTA methods.
• Achieved Dice score improvements of 8.22% in fundus image segmentation.
• Achieved Dice score improvements of 8.15% in heart structure segmentation.

Conclusions:

• SicTTA offers a robust solution for continual test time adaptation in medical image segmentation.
• The proposed CCD, SABE, and SFF components effectively handle domain shifts and limited data.
• SicTTA demonstrates superior performance, enhancing clinical applicability of segmentation models.