Related Experiment Video
Updated: Jun 21, 2026
10:32
Design, Surface Treatment, Cellular Plating, and Culturing of Modular Neuronal Networks Composed of Functionally Inter-connected Circuits
Published on: April 15, 2015
FD-MSP: feature decoupling network with multi-scale prototypes for domain-adaptive polyp segmentation
Wenqi Zhong1, Xu Yang2, Shengyuan Liu3
1Faculty of Applied Sciences, Macao Polytechnic University, Macao SAR, China.
BMC Medical Imaging
|June 19, 2026
Summary
This study introduces FD-MSP, a new method for polyp segmentation in colonoscopy images. It improves deep learning models by decoupling features and using multi-scale prototypes, enhancing colorectal cancer detection across different devices.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Deep learning models excel at polyp segmentation for colorectal cancer detection but struggle with variations across different endoscopic devices and imaging protocols.
- Unsupervised Domain Adaptation (UDA) aims to bridge this domain gap using labeled source data and unlabeled target data.
- Existing UDA methods entangle domain-specific and invariant features and use single-scale representations, failing to capture polyp size variations.
Purpose of the Study:
- To propose FD-MSP, a novel feature decoupling network with multi-scale prototypes for robust cross-domain polyp segmentation.
- To address limitations of existing UDA methods by explicitly decoupling feature alignment and multi-scale prototype learning.
- To improve the accuracy and generalizability of polyp segmentation models in colonoscopy images.
Main Methods:
- FD-MSP employs a dual-stream shallow-encoder to decouple domain-invariant polyp features from equipment-specific variations.
- A multi-scale grouped prototype layer with dilated convolutions captures polyp patterns at various granularities.
- An online pseudo-fusion adapter and a dual-branch gated head adaptively fuse predictions for improved segmentation.
Main Results:
- FD-MSP demonstrated consistent performance improvements over existing UDA methods on three public colonoscopy datasets.
- The proposed method effectively handles domain shifts caused by different imaging protocols and endoscopic devices.
- Feature decoupling and multi-scale prototype learning proved crucial for enhanced cross-domain polyp segmentation.
Conclusions:
- FD-MSP offers a significant advancement in unsupervised domain adaptation for colonoscopy polyp segmentation.
- The proposed architecture effectively addresses the challenges of feature entanglement and scale variation in cross-domain segmentation.
- This work contributes to more reliable early detection of colorectal cancer through improved automated polyp identification.
