Test-time adaptation for surgical instrument segmentation under visual degradations: selective updating and temporal constraints
Related Experiment Video
Updated: Nov 9, 2026
Automated Joint Space Detection Improves Bone Segmentation Accuracy
Published on: November 28, 2025
Shiqi Jiang1,2, Hai Peng1, Shuoxin Gu1
1School of Automation, Chengdu University of Information Technology, Chengdu, China.
Introduction:
Visual perception for robot-assisted minimally invasive surgery is sensitive to low illumination, smoke, blood interference, and imaging blur, which can degrade segmentation performance during continuous video.
Methods:
We propose Lightweight Spatiotemporal Test-Time Adaptation (LiteST-TTA), combining a Top-k predictive-entropy uncertainty gate, motion-compensated historical predictions, hybrid temporal supervision, and updates restricted to Batch Normalization affine parameters. We evaluate five methods on EndoVis 2017 under 12 controlled degradation conditions (four types × three severity levels).
Results:
TENT yields a macro-average DSC of 0.8881 and LiteST-TTA yields 0.8879. A significant degradation-type-by-severity interaction is observed for DSC (Greenhouse-Geisser corrected p = 0.031) and DSC improvement over Source (p < 0.001). Under an identical 24.2% update schedule, LiteST-TTA and Selective Entropy achieve nearly identical DSC (0.886967 vs. 0.886966). Statistics-only BN adaptation yields a DSC of 0.8864. In sequence-disjoint evaluation, LiteST-TTA improves severe low-light performance (0.8083 vs. 0.3251 for Source), whereas Source is superior in 10 of 12 held-out conditions.
Discussion:
Adaptation benefits depend strongly on degradation type, severity, normalization adaptation, update dynamics, and computational cost. These factors should be considered jointly when applying online adaptation to surgical video.