A Multiscale Colorectal Polyp Segmentation Method Based on Frequency-Domain Feature Decoupling and H-Dimensional Dynamic Group Attention

we propose an encoder-decoder framework FDHNet that achieves robust segmentation. The encoder utilizes PVT to extract multi-scale low-level features and incorporates an Frequency-Morphology Feature Decoupling (FMFD) module. By applying Fast Fourier Transform (FFT), features are explicitly decoupled into low-frequency structural components and high-frequency textural components. The low-frequency stream is processed by Mamba to capture long-range morphological dependencies, while the high-frequency stream is enhanced by a differentiable morphological gradient to strengthen edge contrast. Subsequent channel attention-based dynamic fusion significantly mitigates semantic gaps. Furthermore, a 3D Group-Calibrated Attention Module (3D-GCAM) incorporating an H-dimensional attention is designed to dynamically calibrate features along the longitudinal intestinal wall direction, effectively suppressing specular artifacts. The decoder introduces an Mamba Inception (MI) module that combines Mamba with an Inception-based four-branch parallel structure to mine multi-scale features, dynamically weighted by channel attention for comprehensive macro-micro optimization. Experimental evaluations on five public datasets demonstrate that FDHNet achieves mDice/mIoU scores of $0.909/0.846$, $0.941/0.896$, $0.934/0.888$, $0.812/0.736$, and $0.830/0.752$ on EndoScene, ClinicDB, Kvasir-SEG, ETIS-LaribPolypDB, and CVC-ColonDB respectively, outperforming existing methods significantly. This framework provides a feasible new paradigm for precise clinical diagnosis of colorectal polyps.

Usage

Setup

Python 3.8
Pytorch 1.7.1
torchvision 0.8.2

Dataset

Download the training and test datasets and move them into ./dataset/, see Google Drive/Baidu Drive [code:dr1h].

Pre-trained model

Download the pre-trained model from Baidu Drive [code:w4vk], and then put it in the ./pretrained_pth folder for initialization.

Train the model

Clone the repository

git clone https://github.com/baiboat/FDHNet.git
cd FDHNet 
bash train.sh

Test the model

cd FDHNet 
bash test.sh

Evaluate the trained model

cd FDHNet 
python Eval.py

Well-trained model

Baidu Drive [code:1228] and put the model in directory ./model_pth.

License

The source code is free for research and education use only. Any commercial use should get formal permission first.

Acknowledgement

Thanks PraNet, EAGRNet, MSEG and Polyp-PVT for serving as building blocks of FDHNet.