A Pytorch general Medical image Segmentation Framework

The main contributors to the project are Chen Yunqing(myself) and Chen Cheng.

This repository is a general Medical image Segmentation Framework and highly based on https://github.com/MontaEllis/Pytorch-Medical-Segmentation

中文说明: ZH_README

---

The model have completed

If the following model has been helpful to you, we hope you can cite the corresponding reference.

Publication Date	Model Name	The First and Last Authors	Title	Reference
2016-10	3D U-Net	Özgün Çiçek and Ronneberger, Olaf	3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation	MICCAI2016
2016-10	3D V-Net	Fausto Milletari and Seyed-Ahmad Ahmadi	V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation	3DV2016
2017-09	3D DenseVoxelNet	Lequan Yu and Jing Qin & Pheng-Ann Heng	Automatic 3D Cardiovascular MR Segmentation with Densely-Connected Volumetric ConvNets	MICCAI2017
2017-09	3D DenseNet	Toan Duc Bui and Taesup Moon	3D Densely Convolutional Networks for Volumetric Segmentation	arxiv
2017-05	3D HighResNet	Wenqi Li and M. Jorge Cardoso & Tom Vercauteren	On the Compactness, Efficiency, and Representation of 3D Convolutional Networks: Brain Parcellation as a Pretext Task	IPMI2017
2017-05	3D Residual U-Net	Kisuk Lee and H. Sebastian Seung	Superhuman Accuracy on the SNEMI3D Connectomics Challenge	arxiv
2021-10	CSR-Net	Cheng Chen and Ruoxiu Xiao	CSR-Net: Cross-Scale Residual Network for multi-objective scaphoid fracture segmentation	CIBM2021
2022	UNETR	Ali Hatamizadeh and Daguang Xu	UNETR: Transformers for 3D Medical Image Segmentation	CVPR2022
2023	IS	Chen Cheng and Ruoxiu Xiao	Integration- and separation-aware adversarial model for cerebrovascular segmentation from TOF-MRA	CMPB

Usage

Requirements

The recommend python and package version:

• python 3.10.0
• pytorch 1.13.1 (do not use torch2.0, it will cause program failed to start)

Train

here we use an example(Traning 3D Unet) to teach you how use this repository

python train.py config=unet

To specify the folder name for the current save, you can modify the corresponding parameter using config.XXX=XXX.

python train.py config=unet config.name=unet-1

All files during the model training process will be saved in ./logs/unet-1

all parameter can be setted in conf/unet.yaml

Global configuration

Considering that many settings are common to all configuration files, such as data_path, num_epochs, etc., to avoid repetitive work, we have placed these common parameters in conf/config.yaml. All configuration files will have these properties from config.yaml.

Taking num_workers (defaulted to 18 in config.yaml) as an example, the priority of parameter overriding is as follows: Command line argument config.num_workers=20 > num_workers=18 in data_3d.yaml > Default value num_workers=18 in config.yaml.

File Structure

Traning logs will be saved like this:

./logs/ying_tof (Corresponding saved folder: ./logs/config.name)
└── 2023-11-24 (Date: year-month-day)
    └── 17-05-02 (Time: hour-minute-second)
        ├── .hydra (Configuration save files)
        │   ├── config.yaml (Configuration for this running)
        │   ├── hydra.yaml
        │   └── overrides.yaml
        ├── train_3d.log (Log during training)
        └── train_tensorboard (Tensorboard folder)
            └── events.out.tfevents.1700816703.medai.115245.0 (file for Tensorboard)

Predict

run the code

python predict.py config=unet config.ckpt=XXX

WARNING: ckpt must be the absolute path of the model, not the relative path

./results/ (Root folder for results)
└── unet (Model name: unet)
    └── 2023-12-04 (Date: year-month-day)
        └── 17-39-30 (Time: hour-minute-second)
            ├── metrics.csv (CSV file containing metrics)
            ├── pred_file (Folder for prediction files)
            │   ├── pred-0000.mhd (Prediction file 0 in MHD format)
            │   ├── pred-0000.zraw (Prediction file 0 in ZRAW format)
            │   ├── pred-0001.mhd (Prediction file 1 in MHD format)
            │   ├── pred-0001.zraw (Prediction file 1 in ZRAW format)
            │   ├── pred-0002.mhd (Prediction file 2 in MHD format)
            │   └── ... (Additional prediction files)
            └── predict.log (Log file for prediction)

Create your own configuration

create new file in path /conf/config, file name ends with .yaml.

For example, if you wanna use 3D Vnet, you can create vnet.yaml, and then set all parameters you wanna set.

how to use the parameters in yaml file

config.PARAM: replace PARAM with the parameters you wanna use

Modify train.py and predict.py to find your model

in train.py, add these codes after the last model

elif config.network == 'NETWORK':
    from models.three_d.NETWORK import NETWORK
    model = NETWORK()

NETWORK: means the network you wanna use

Related Works

if this code is helpful for you, you can cite these for us. Thanks！

 C Chen and R Xiao. Generative consistency for semi-supervised cerebrovascular segmentation from TOF-MRA, 2022, IEEE Transactions on Medical Imaging 42 (2), 346-353
 C Chen and R Xiao. All answers are in the images: A review of deep learning for cerebrovascular segmentation, 2023, Computerized Medical Imaging and Graphics, 102229
 C Chen and J Zhao. Understanding the brain with attention: A survey of transformers in brain sciences, 2023, Brain-X, 1 (3): e29.