CSCA U-Net: A Channel and Space Compound Attention CNN for Medical Image Segmentation

---

Table of Contents

• CSCA U-Net: A Channel and Space Compound Attention CNN for Medical Image Segmentation
  • Table of Contents
  • 1. Overview
    • Papers
    • Architecture
    • Abstract
  • 2. Data Files
    • 2.1 Datasets
    • 2.2 Trained weights
    • 2.3 Predicted Map
  • 3. How to run
    • 3.1 Create Environment
    • 3.2 Training
    • 3.3 Testing and Save Pictures
    • 3.4 Evaluating
  • 4. Result
  • 5. Citation
  • 6. Acknowledge
  • 7. Contact

Table of contents generated with markdown-toc

---

You also can read Chinese version

Please run to get the code:

git clone --recursive https://github.com/xiaolanshu/CSCA-U-Net.git

1. Overview

Papers

This repository provides code for our paper "CSCA U-Net: A Channel and Space Compound Attention CNN for Medical Image Segmentation" accepted for Publication at Artificial Intelligence in Medicine.(papers)

Architecture

Abstract

Image segmentation is one of the vital steps in medical image analysis. A large number of methods based on convolutional neural networks have emerged, which can extract abstract features from multiple-modality medical images, learn valuable information that is difficult to recognize by humans, and obtain more reliable results than traditional image segmentation approaches. U-Net, due to its simple structure and excellent performance, is widely used in medical image segmentation. In this paper, to further improve the performance of U-Net, we propose a channel and space compound attention (CSCA) convolutional neural network, abbreviated CSCA U-Net, which increases the network depth and employs a double squeeze-and-excitation (DSE) block in the bottleneck layer to enhance feature extraction and obtain more high-level semantic features. Moreover, the characteristics of the proposed method are three-fold: (1) channel and space compound attention (CSCA) block, (2) cross-layer feature fusion (CLFF), and (3) deep supervision (DS). Extensive experiments on several available medical image datasets, including Kvasir-SEG, CVC-ClinicDB, CVC-ColonDB, ETIS, CVC-T, 2018 Data Science Bowl (2018 DSB), ISIC 2018 Challenge, and JSUAH-Cerebellum, show that CSCA U-Net achieves competitive results and significantly improves the generalization performance.

2. Data Files

2.1 Datasets

In this subsection, we provide the public data set used in the paper:

• Polyp Datasets (include Kvasir-SEG, CVC-ClinicDB, CVC-ColonDB, ETIS, and CVC-300.) [From PraNet]:
  • Total: [Aliyun], [Baidu]
  • TrainDataset: [Google Drive]
  • TestDataset: [Google Drive]
• 2018 Data Science Bowl: [Aliyun], [Baidu], [Google Drive]
• ISIC 2018 (original from [kaggle]. I converted the images in .tiff format to .png format): [Aliyun], [Baidu], [Google Drive]

2.2 Trained weights

[Aliyun], [Baidu], [Google Drive]

2.3 Predicted Map

[Aliyun], [Baidu], [Google Drive]

3. How to run

3.1 Create Environment

First of all, you need to have a pytorch environment, I use pytorch 1.10, but it should be possible to use a lower version, so you can decide for yourself.

You can also create a virtual environment using the following command (note: this virtual environment is named pytorch, if you already have a virtual environment with this name on your system, you will need to change environment.yml manually).

conda env create -f docs/enviroment.yml

3.2 Training

You can run the following command directly:

sh run.sh ### use stepLR
sh run_cos.sh ### use CosineAnnealingLR 

If you only want to run a single dataset, you can comment out the irrelevant parts of the sh file, or just type something like the following command from the command line:

python Train.py --model_name CSCAUNet --epoch 121 --batchsize 16 --trainsize 352 --train_save CSCAUNet_Kvasir_1e4_bs16_e120_s352 --lr 0.0001 --train_path $dir/data/TrainDataset --test_path $dir/data/TestDataset/Kvasir/  # you need replace ur truely Datapath to $dir.

3.3 Get all predicted results (.png)

If you use a sh file for training, it will be tested after the training is complete.

If you use the python command for training, you can also comment out the training part of the sh file, or just type something like the following command at the command line:

python Test.py --train_save CSCAUNet_Kvasir_1e4_bs16_e120_s352 --testsize 352 --test_path $dir/data/TestDataset

3.4 Evaluating

• For evaluating the polyp dataset, you can use the matlab code in eval or use the evaluation code provided by [UACANet].
• For other datasets, you can use the code in evaldata.
• The reason for using a different evaluation code is to use the same methodology in the evaluation as other papers that did experiments on the dataset.

4. Result

• Polyp Datasets

  

• 2018 DSB

• ISIC 2018

• JSUAH-Cerebellum

• Qualitative Results

  • Polyp Segmentation
  • 2018 DSB
  • ISIC 2018
  • Fetal Ultrasound （Private）

5. Citation

Please cite our paper if you find the work useful:

@article{shu2024102800,
title = {CSCA U-Net: A channel and space compound attention CNN for medical image segmentation},
journal = {Artificial Intelligence in Medicine},
volume = {150},
pages = {102800},
year = {2024},
issn = {0933-3657},
doi = {https://doi.org/10.1016/j.artmed.2024.102800},
url = {https://www.sciencedirect.com/science/article/pii/S0933365724000423},
author = {Xin Shu and Jiashu Wang and Aoping Zhang and Jinlong Shi and Xiao-Jun Wu},
keywords = {U-net, Channel and spatial compound attention, Cross-layer feature fusion, Deep supervision, Medical image segmentation}
}

6. Acknowledge

• Many of the training strategies, datasets, and evaluation methods in this paper are based on PraNet. I admire the open source spirit of Dr. Deng-Ping Fan and other authors, and am very grateful for the help provided by this work on PraNet.

7. Contact

• 📫 reach me for email:[email protected]

• reach me for QQ: 872845991