EViT-Unet: U-Net Like Efficient Vision Transformer for Medical Image Segmentation on Mobile and Edge Devices

Xin Li, Wenhui Zhu, Xuanzhao Dong, Oana M. Dumitrascu, Yalin Wang

Code for the paper: arXiv

Abstract

With the rapid development of deep learning, CNN-based U-shaped networks have succeeded in medical image segmentation and are widely applied for various tasks. However, their limitations in capturing global features hinder their performance in complex segmentation tasks. The rise of Vision Transformer (ViT) has effectively compensated for this deficiency of CNNs and promoted the application of ViT-based U-networks in medical image segmentation. However, the high computational demands of ViT make it unsuitable for many medical devices and mobile platforms with limited resources, restricting its deployment on resource-constrained and edge devices. To address this, we propose EViT-UNet, an efficient ViT-based segmentation network that reduces computational complexity while maintaining accuracy, making it ideal for resource-constrained medical devices. EViT-UNet is built on a U-shaped architecture, comprising an encoder, decoder, bottleneck layer, and skip connections, combining convolutional operations with self-attention mechanisms to optimize efficiency. Experimental results demonstrate that EViT-UNet achieves high accuracy in medical image segmentation while significantly reducing computational complexity.

Comparison of model size, FLOPs, and performance (Dice on Synapse dataset). The area of each circle is propor- tional to the number of parameters (model size).

Prerequisites

1. Clone the repo to your local:

git clone https://github.com/Retinal-Research/EVIT-UNET.git

2. We use Python 3.10.15 and Pytorch 2.0.1. Build the virtual enviroment with conda

conda env create -f environment.yml

Data preparation

Synapse dataset is used in our paper. Download

Weight Download

The model weight can be downloaded from Here

Train

• Set your own parameters, and start the train process, eg.:

python train_unetKD.py --root_path /data/Synapse --output_dir ./results --batch_size 32 --max_epochs 300 --img_size 224

Test

python test_unet.py

Visualization

Thanks for the code provided by:

EfficientFormerV2 https://github.com/snap-research/EfficientFormer/blob/main/README.md

Swin-Unet https://github.com/HuCaoFighting/Swin-Unet

SelfReg-UNet https://github.com/ChongQingNoSubway/SelfReg-UNet

UCTransNet https://github.com/McGregorWwww/UCTransNet/blob/main/nets/UCTransNet.py