Skip to content
Open access

Frequency‐Guided Cascaded Refinement for Low‐Contrast Colorectal Polyp Segmentation

Nov 2026 · International journal of imaging systems and technology (Print) · 0 citations · 20 references

Abstract

Accurate polyp segmentation plays an important role in computer‐aided colonoscopy and colorectal cancer screening. However, weak lesion contrast, specular highlights, intestinal folds, residual fluid, and mucosal texture make it difficult to distinguish polyps from surrounding mucosa. These factors can produce prominent local frequency responses, but such responses do not always correspond to lesion regions. To address this issue, we propose a Frequency‐guided Cascaded Refinement Network (FCRNet) for colonoscopic polyp segmentation. FCRNet incorporates a Frequency Boundary Prior (FBP) module to extract directional high‐frequency structural cues from the input image through Haar wavelet decomposition. Rather than directly treating these cues as lesion predictions, a Semantic Refinement Module (SRM) integrates them with top‐down semantic feedback to suppress semantically inconsistent activations and progressively refine decoder representations. Extensive experiments on five public polyp segmentation benchmarks demonstrate the effectiveness of the proposed method. FCRNet achieves mDice/mIoU scores of 93.3%/89.1% on Kvasir‐SEG and 94.8%/90.8% on CVC‐ClinicDB, while maintaining strong cross‐dataset performance on CVC‐ColonDB, ETIS, and CVC‐300. Ablation studies further confirm that FBP and SRM provide complementary contributions. These results indicate that semantically guided use of frequency cues can improve polyp segmentation under challenging imaging conditions.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.