What question did this study set out to answer?

The work addresses challenges in polyp segmentation during colorectal cancer screening.

May 6, 2026Biology0 citationsOpen Access

WCEDSAM: A Lightweight Multi-Scale Colonoscopy Polyp-Segmentation Network Combining Frequency-Domain Decomposition and Adaptive Feature Enhancement

LWLei WangZhejiang Chinese Medical University TWTongyu WangZhejiang Chinese Medical University SLSitong LiuZhejiang Normal University

Key Points

The work addresses challenges in polyp segmentation during colorectal cancer screening.
Developed a lightweight medical image segmentation model called WCEDSAM.
Incorporated a Wavelet Transform-based component to extract overlapping features.
Utilized a DSConv-ECA module to efficiently capture local features and reduce parameter count.
WCEDSAM achieved mean Dice scores of 0.9383 on Kvasir-SEG and 0.9376 on CVC-ClinicDB.
Cross-domain evaluations showed mDice scores of 0.9189 on CVC-ColonDB and 0.8961 on CVC-300.
Substantially outperformed methods like UNet++ and TransUNet.

Abstract

Colorectal cancer screening is challenged by variations in polyp morphology, indistinct polyp boundaries, and the high computational costs associated with current models. To address these issues, a lightweight medical image segmentation model, WCEDSAM, has been developed. WCEDSAM is based on a modified, compact version of MedSAM, which incorporates a Wavelet Transform-based component to extract and separate overlapping features at the pixel level. Additionally, a DSConv-ECA module is positioned before the ViT encoder to capture local features efficiently while reducing parameter count and enhancing inter-channel communication. Experimental results demonstrate that WCEDSAM achieves top performance on five public datasets, including Kvasir-SEG and CVC-ClinicDB, with 15.38 million parameters, achieving mean Dice (mDice) scores of 0.9383 on Kvasir-SEG and 0.9376 on CVC-ClinicDB. Cross-domain evaluations yield mDice scores of 0.9189 on CVC-ColonDB, 0.8961 on CVC-300, and 0.7765 on ETIS datasets, respectively, substantially outperforming other methods such as UNet++ and TransUNet.

问 AI

Bookmark

View Full Paper