GEN-Guard: correcting generalization failures for deployable federated surgical AI

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Federated Learning (FL) enables collaborative AI model training on sensitive surgical data without direct sharing.
• Standard FL evaluation can lead to suboptimal deployment due to performance leakage, where models overfit internal data and fail to generalize.
• This undermines the goal of robust real-world generalization in surgical AI.

Purpose of the Study:

• To identify and address the critical failure mode of performance leakage in federated surgical AI.
• To propose a practical framework, GEN-Guard, for detecting and correcting generalization failures in FL models.
• To enhance the cross-institutional robustness and zero-shot adaptation capabilities of surgical AI.

Main Methods:

• Introduction of GEN-Guard, a post-hoc framework for federated surgical AI.
• Integration of Generalization Detection via Client-Blocked Evaluation (CBE) to prevent performance leakage.
• Implementation of Generalization Correction through Disagreement-Aware Distillation (DAD) for adaptive feature-level corrections.

Main Results:

• Quantification of performance leakage, revealing Model Selection Failures (MSFs) exceeding 80% under standard evaluation.
• GEN-Guard demonstrated effectiveness in surgical phase recognition and polyp segmentation tasks.
• Consistent improvements observed: up to 2 points in-federation F1 scores, up to 3 points in unseen-institution performance, and 3-9 points in worst-case institutional performance.

Conclusions:

• Performance leakage is a significant, under-recognized risk in federated surgical AI.
• GEN-Guard offers a privacy-preserving, practical solution to detect and correct generalization failures post-training.
• The framework enhances FL reliability for real-world surgical applications by improving robustness and zero-shot generalization.