Multi-View Test-Time Adaptation for Semantic Segmentation in Clinical Cataract Surgery.

Area of Science:

• Ophthalmology
• Computer Vision
• Medical Imaging

Background:

• Cataract surgery utilizes semantic segmentation for computer-assisted interventions.
• Variations in clinical settings cause domain shifts, impacting segmentation model performance.
• Existing domain adaptation methods raise privacy concerns due to data centralization.

Purpose of the Study:

• To propose a novel algorithm, Multi-view Test-time Adaptation (MUTA), for robust cataract surgical scene segmentation.
• To address domain shifts and privacy concerns in cross-center clinical applications.
• To enable effective model adaptation during inference without data centralization.

Main Methods:

• MUTA employs multi-view learning for enhanced model training in the source domain.
• Multi-view decoders are integrated during the training phase to improve robustness.
• Test-time adaptation with multi-view knowledge distillation is used for inference-phase adaptation.

Main Results:

• MUTA effectively learns a robust source model for cataract surgery segmentation.
• The algorithm successfully adapts the model to target data during practical inference.
• Experiments in simulated cross-center scenarios validated MUTA's effectiveness.

Conclusions:

• MUTA offers a privacy-preserving solution for domain adaptation in cataract surgery segmentation.
• The algorithm enhances the reliability of computer-assisted interventions across different clinical environments.
• MUTA facilitates on-site model updates without compromising patient data privacy.