Personalizing Federated Instrument Segmentation With Visual Trait Priors in Robotic Surgery

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Personalized federated learning (PFL) enables collaborative model training across multiple clinical sites while preserving data privacy.
• Existing PFL methods for surgical instrument segmentation (SIS) often overlook personalization in attention mechanisms and fail to address appearance diversity and instrument shape similarity inherent in surgical data.
• Improving the accuracy and robustness of SIS is crucial for advancing surgical robotics and computer-assisted surgery.

Purpose of the Study:

• To introduce PFedSIS, a novel PFL method designed to enhance SIS performance by incorporating visual trait priors.
• To address the limitations of existing PFL methods by focusing on personalization of multi-headed self-attention and accounting for appearance and shape variations.
• To improve the accuracy and reliability of surgical instrument segmentation across diverse clinical settings.

Main Methods:

• PFedSIS integrates three key components: global-personalized disentanglement (GPD) for head-wise attention personalization, appearance-regulation personalized enhancement (APE) for customized layer-wise aggregation using hypernetworks, and shape-similarity global enhancement (SGE) for maintaining and sharing mutual instrument shape information.
• GPD facilitates the first head-wise assignment for multi-headed self-attention personalization.
• APE preserves unique site-specific appearance representations and leverages inter-site differences through customized, site-specific parameter aggregation.
• SGE enhances cross-style shape consistency and utilizes site-specific shape contributions to update global parameters.

Main Results:

• PFedSIS achieved superior performance compared to state-of-the-art methods in surgical instrument segmentation.
• Specific performance gains include +1.51% Dice, +2.11% IoU, -2.79 in Average Symmetric Surface Distance (ASSD), and -15.55 in Hausdorff Distance 95% (HD95).
• The proposed method effectively handles appearance diversity and instrument shape similarity, leading to significant improvements in segmentation accuracy.

Conclusions:

• PFedSIS represents a significant advancement in personalized federated learning for surgical instrument segmentation.
• The method's novel components effectively address personalization challenges, appearance diversity, and shape similarity, leading to enhanced segmentation performance.
• PFedSIS demonstrates the potential to improve the reliability and applicability of AI-driven surgical tools in real-world clinical environments.