PrysmNet: Enhancing Cross-Domain Polyp Segmentation with Salience and Multimodal Guidance

Overview

PrysmNet introduces a novel biologically inspired salience module and advanced training strategies to improve polyp segmentation accuracy and generalization across diverse clinical domains. It addresses challenges in detecting small and boundary-challenged polyps, achieving sharper segmentation masks and robust performance on multi-center datasets.

Background

Colorectal cancer (CRC) often develops from benign polyps, making early detection and removal critical to reducing mortality. Colonoscopy is the gold standard for screening but is operator-dependent and challenged by the diversity of polyp appearances, leading to missed lesions. Automated polyp segmentation can assist clinicians by providing precise lesion boundaries, yet existing deep learning models often fail to generalize well across different clinical settings due to domain shifts. There is a pressing need for methods that maintain high accuracy on unseen datasets and handle difficult cases such as diminutive or faint polyps.

Data Highlights

Key Findings

• PrysmNet's biologically inspired salience module (BSM) dynamically enhances lesion boundary features, resulting in sharper and more accurate segmentation masks.
• The multi-level foundation model distillation module (FMDM) effectively transfers knowledge from foundation models to improve feature representation.
• The multi-modal guidance module (MGM) leverages auxiliary structural and textural information to create invariant features, enhancing robustness to domain shifts.
• Existing models show high in-domain Dice scores (0.85–0.95) but suffer performance drops (to ~0.6–0.7 Dice) on unseen multi-center datasets like PolypGen.
• PrysmNet demonstrates improved generalization across diverse imaging conditions and patient populations, addressing limitations of prior single-center trained models.

Clinical Implications

PrysmNet's enhanced segmentation accuracy and cross-domain robustness can assist endoscopists in reliably identifying and delineating polyps, including challenging small or flat lesions. This may reduce missed polyps and improve complete resection rates, ultimately contributing to better colorectal cancer prevention. The system's ability to generalize across centers supports broader clinical deployment without extensive retraining.

Conclusion

PrysmNet represents a significant advancement in automated polyp segmentation by combining biologically inspired architectural innovations with sophisticated training strategies to achieve consistent, high-quality segmentation across diverse clinical domains. This approach addresses critical gaps in real-world applicability and paves the way for more reliable AI-assisted colonoscopy.

References

1. Ali et al. 2022 -- PolypGen: A Multi-Center Polyp Segmentation Dataset
2. Bernal et al. 2017 -- Benchmark for Polyp Detection and Segmentation
3. MICCAI EndoScene and GIANA Challenges -- Polyp Segmentation Competitions
4. CVC-ClinicDB Dataset -- Early Polyp Segmentation Benchmark
5. Kvasir-SEG Dataset -- Polyp Segmentation Dataset