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

Abstract

Test Time Adaptation (TTA) enhances model robustness by adapting to unseen domains during testing. Existing methods typically rely on large batch sizes or pseudo-label generation, which are often impractical in clinical settings where data is limited and subject to continual distribution shifts. Moreover, pseudo-labels can fail to align with the source domain distribution, leading to unreliable results. To address these challenges, we propose SicTTA, a novel approach for single-image continual test time adaptation in medical image segmentation. SicTTA introduces Class Compact Density (CCD) analysis to estimate uncertainty and select Source-Friendly Target (SFT) images that align closely with the source model's knowledge. We maintain an SFT image and feature pool, managed with a first-in-first-out strategy to handle size constraints. For each non-SFT image, Source-Aligned Batch Enhancement (SABE) selects the top K images based on latent feature similarity, creating an enhanced batch that improves statistics normalization. Additionally, Similarity-driven Feature Fusion (SFF) aligns the test image with the enhanced batch's distribution, preserving crucial features. SicTTA outperforms seven state-of-the-art TTA methods in fundus image and heart structure segmentation across a sequence of target domains, achieving Dice score improvements of 8.22 and 8.15 percentage points over the source model, respectively.