PRISM: Past-Regularized Iterative Self-Distillation With Momentum for Polyp Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Colorectal cancer is a significant health concern, with early detection crucial for effective treatment.
• Accurate segmentation of colon polyps in medical images is vital for early diagnosis.
• Current deep learning methods for polyp segmentation require extensive data and computational resources, limiting their clinical applicability.

Purpose of the Study:

• To introduce PRISM, a novel momentum-based self-distillation method for enhancing colon polyp segmentation.
• To improve segmentation performance without increasing inference costs.
• To develop a generalizable solution for polyp segmentation across diverse clinical settings.

Main Methods:

• PRISM utilizes a momentum-based self-distillation approach.
• A temporally smoothed teacher model is created using exponential moving average (EMA) on model weights during training.
• The EMA-based teacher provides stable, adaptive supervision signals that co-evolve with the student model.

Main Results:

• PRISM achieved a Dice score of 0.81 and an IoU of 0.75 on colonoscopy datasets from five medical centers.
• The method demonstrated superior performance compared to baseline and conventional self-distillation techniques.
• Validation on an independent dataset confirmed PRISM's generalization capabilities.

Conclusions:

• PRISM offers a computationally efficient and generalizable solution for colon polyp segmentation.
• The EMA-based teacher model effectively enhances segmentation accuracy.
• This method holds promise for improving early detection and treatment of colorectal cancer.