AttentionBased-MIS

Progressive Attention Module for Segmentation of Volumetric Medical Images

By Minghui Zhang, Hong Pan, Yaping Zhu, and Yun Gu

Institute of Medical Robotics, Shanghai Jiao Tong University

Department of Computer Science and Software Engineering, Swinburne University of Technology

Medical Image Segmentation Based on Attention mechanism(Fine-Tune for natural image[semantic/instance] segmentation).

This project is dedicated to

Collecting and re-implementing basic models and different attention mechanisms, transforming them modular and portable.
Proposing a novel attention mechanism ———— 'Slice Attention' and 'Progressive Attention Module' tailored for 3D data Segmentation.

Main purpose is used in 3D Medical Image Segmentation. Fine-tune for Other CV tasks need attention is easily meanwhile.

dataset

We have used three medical image dataset,both 3D MRI format.You can use other formats like CT only need to constrcuct a concrete datatset loader in /data.

Dataset	Link
BraTS 2018 dataset	link
MALC dataset	link
HVSMR dataset	link

models

This section include basic model(for segmentation or feature extraction) and different attention mechanisms.Each attention mechanism can recalibrate multi-dim feature maps across their own functional domain.

Most attention mechanisms can be modularized and integrated into any sub feature maps(e.g. each encoder in 3D UNet or each block in VNet) if not special noted.

All models and basic modules are in /models,We provide the pre-trained Models in Model_zoo.md:

Here,we integrate and provide brief description for basic segmentation models and attention modules used in our experiments. Most of them have conducted and you can find pretrained models in Model_zoo.md. Some modules have not finished and we will keep conducting experiments and updating this project.

Basic Segmentation models tailored for Medical Images(Both 2D and 3D)

Module	Paper	Name
3D UNet	link	3D UNet
VNet	link	VNet
U-Net	link	2D UNet
DeepMedic	link	DeepMedic
VoxResNet	link	VoxResNet
H-DenseUNet	link	H-DenseUNet

Attention Module

Module	Paper	Name
Squeeze-and-Excitation	link	SE
Convolutional Block Attention Module	link	CBAM
Project&Excitation	link	PENet
Attention U-Net	link	AG
AnatomyNet	link	AnatomyNet
Progressive Attention Module	link	PANet
Class Activation Map	link	CAM
Spatial Transformer Net	link	STN
Split Attention	link	SpA

Train and Test

To fully understand the architecture of our projects, please take a reference to

Visualization of the projects arch

The Technical Report has clearly demonstrate the procedure of training and test, including loading data, pre-processing data, loading pretrained models, training and testing models, saving model weights and other log info, etc.

Train Demo

You can use the following command to train a model, demo_train.py is a script with detailed annotations.

python demo_train.py --dataroot $DATASET_DIR \
                     --name $EXPERIMENT_NAME \
                     --checkpoints_dir $MODEL_SAVEDIR \
                     --model $MODEL \
                     --dataset_mode $DATASET_MODE \
                     --in_channels $INPUT_CH \
                     --out_channels $OUTPUT_CH \
                     --gpu_ids $GPU_IDS

Test Demo

You can use the following command to test a model, demo_test.py is a script with detailed annotations.

python demo_test.py  --dataroot $DATASET_DIR \
                     --name $EXPERIMENT_NAME \
                     --checkpoints_dir $MODEL_LOADDIR \
                     --model $MODEL \
                     --dataset_mode $DATASET_MODE \
                     --in_channels $INPUT_CH \
                     --out_channels $OUTPUT_CH \
                     --gpu_ids $GPU_IDS

Environment

The research in this study is carried out on both Windows10 and Ubuntu16.04 with 4 NVIDIA 1080Ti GPU with 44GB memory.

The deep learning framework is PyTorch≥1.1.0 and Torchvision ≥0.4.0.

Some python libraries are also necessary, you can use the following command to set up.

pip install -r requirements.txt

📝 Citation

If you find this repository or our paper useful, please consider citing our paper:

@article{zhang2022progressive,
  title={Progressive attention module for segmentation of volumetric medical images},
  author={Zhang, Minghui and Pan, Hong and Zhu, Yaping and Gu, Yun},
  journal={Medical Physics},
  volume={49},
  number={1},
  pages={295--308},
  year={2022},
  publisher={Wiley Online Library}
}

Puzzled-Hui/AttentionBased-MIS