CVPixUNet

Rethinking U-Shape Segmentation Network: Towards CNN- & ViT-based Hybrid Network with Dynamic Adaptive Pixel-Level Feature Learning for Retinal Vessel Segmentation

Motivation

Exploring Pixel-Level DynamiC CNN + Pixel-Level ViT for medical image segmentation.

Requirements

• Pytorch
• Some basic python packages such as Numpy, Scikit-image, SimpleITK, Scipy ......

DataSets

DRIVE CHASEDB1

Results

Segmentation Visualization of DRIVE Dataset Including: (a) Original Image; (b) Ground Truth (c) UNet (d) UNet++ (e) DeepLabv3++ (f) ConvUNext (g) DSCNet and (h) CVPixUNet.

Segmentation Visualization of CHASEDB1 dataset, where (a) Original image (b) Ground truth (c) UNet (d) UNet++ (e) DeepLabv3+ (f) ConvUNeXt (g)DSCNet and (h) CVPixUNet

Usage

1. Clone the repo:

git clone https://github.com/ziyangwang007/CVPixUNet.git
cd CVPixUNet

2. Train the model

python train.py 

3. Test the model

python val.py 

Reference

Ziyang Wang, Mian Wu. "Rethinking Hybrid U-Shape Network with Pixel-Level Feature Learning for Retinal Vessel Segmentation." IEEE Access (2026).