Towards Clinically Applicable Large-Model-Based Privacy-Preserving Polyp Segmentation: A Federated LoRA Approach to...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate colonoscopy polyp segmentation is crucial for detecting lesions and optimizing clinical workflows.
• Deploying large AI models in healthcare is challenging due to privacy concerns, computational demands, and data variability.

Purpose of the Study:

• To introduce PolypSAMFL, a novel framework for privacy-preserving, high-precision polyp segmentation.
• To integrate low-rank adaptation (LoRA) with the Segment Anything Model (SAM) within a federated learning (FL) approach.

Main Methods:

• Utilized federated learning (FL) with Segment Anything Model (SAM) and low-rank adaptation (LoRA) to enable privacy-preserving model training on distributed datasets.
• Froze most SAM parameters, finetuning only compact LoRA modules to reduce communication overhead.
• Implemented a boundary-aware loss function and multiresolution mask synthesis for improved polyp boundary delineation.

Main Results:

• Achieved a mean Dice score of 0.987 and intersection-over-union (IoU) of 0.976 on four public colonoscopy datasets.
• Demonstrated superior performance compared to state-of-the-art methods while maintaining data locality.
• Significantly reduced communication costs associated with federated learning.

Conclusions:

• PolypSAMFL offers a scalable, privacy-preserving solution for AI-driven colonoscopy workflows.
• The framework enhances polyp segmentation accuracy and reliability, aligning with healthcare privacy regulations and resource constraints.
• Validated clinical utility for real-world AI applications in gastrointestinal endoscopy.