An efficient fine tuning strategy of segment anything model for polyp segmentation

Area of Science:

• Medical imaging
• Artificial intelligence in healthcare
• Gastroenterology

Background:

• Colon cancer is a global health concern, necessitating accurate and cost-effective detection methods.
• Colon polyp segmentation is crucial for early diagnosis, but faces challenges like high annotation costs and limited accuracy.
• Existing fine-tuning methods for models like Segment Anything Model (SAM) can lead to catastrophic forgetting, hindering generalization.

Purpose of the Study:

• To propose and evaluate an efficient fine-tuning strategy for SAM to improve colon polyp segmentation.
• To address the problem of catastrophic forgetting in SAM during fine-tuning for medical imaging tasks.
• To enhance the performance of colon polyp segmentation, particularly in few-shot learning scenarios.

Main Methods:

• Developed PSF-SAM, an efficient fine-tuning method for SAM by freezing most parameters and optimizing specific structures.
• Applied PSF-SAM to colon polyp segmentation tasks.
• Evaluated performance using metrics such as mDice coefficients and mIoU on Kvasir-SEG and CVC-ClinicDB datasets.

Main Results:

• PSF-SAM demonstrated superior performance compared to conventional fine-tuning methods on colon polyp segmentation.
• The proposed method significantly improved accuracy, as indicated by higher mDice and mIoU scores.
• PSF-SAM showed notable advantages in few-shot learning scenarios, outperforming existing techniques.

Conclusions:

• PSF-SAM effectively mitigates catastrophic forgetting and enhances colon polyp segmentation accuracy.
• The proposed efficient fine-tuning strategy offers a promising solution for improving colon cancer detection through medical image analysis.
• PSF-SAM presents a valuable advancement for few-shot learning in medical segmentation tasks.