PrysmNet a polyp refining system using salience and multimodal guidance for reproducible cross domain segmentation.

Abstract

Colorectal cancer prevention benefits from accurate and reproducible polyp segmentation, yet cross-domain generalization and boundary precision remain challenging in real-world deployments. We propose Prysm-Net, a ViT-based framework designed to address these issues through architectural innovation and advanced training guidance. Our model is augmented with a biologically inspired salience module (BSM) that dynamically sharpens boundary-relevant features. To further enhance robustness without increasing inference costs, we introduce two training-only strategies: (i) foundation-model distillation from SAM, which transfers knowledge at the output, boundary, and feature levels, and (ii) multi-modal guidance that injects auxiliary structural and textural cues via gated cross-attention. Extensive experiments on standard in-domain benchmarks and challenging cross-domain datasets demonstrate that Prysm-Net achieves superior segmentation accuracy and robust generalization compared to state-of-the-art methods, all while maintaining a lightweight inference process by disabling auxiliary guidance at test time.