Polyp Segmentation Using Wavelet-Based Cross-Band Integration for Enhanced Boundary Representation

Researchers Haesung Oh and Jaesung Lee have introduced a novel AI model for polyp segmentation that significantly improves boundary detection in colonoscopy images. Presented in a paper at the NeurIPS 2025 MedEurIPS workshop, the model tackles a core challenge in computer-aided diagnosis: accurately localizing polyp boundaries despite low mucosal contrast, uneven illumination, and color similarity with surrounding tissue. The key innovation stems from a quantitative analysis revealing that grayscale representations in the wavelet domain consistently preserve higher boundary contrast than standard RGB images across all frequency bands. This finding challenged conventional RGB-only approaches and formed the foundation for their new architecture.

The proposed model integrates grayscale and RGB representations through a complementary frequency-consistent interaction mechanism. This wavelet-based cross-band integration allows the AI to enhance boundary precision while maintaining the structural coherence of the polyp. Extensive validation on four established medical imaging benchmarks demonstrated that the approach achieves superior robustness and precision compared to existing models. For medical professionals, this translates to a more reliable AI assistant during colonoscopies, potentially reducing missed diagnoses and improving early detection rates for colorectal cancer, a leading cause of cancer mortality worldwide. The next steps likely involve clinical validation and integration into real-time endoscopic systems.