momopusheen/A3D_SATr

Code of SATr: Slice Attention with Transformer for Universal Lesion Detection

Code for paper "SATr: Slice Attention with Transformer for Universal Lesion Detection"

This repository contains the state-of-the-art version (A3D-SATr) of our paper SATr (MICCAI'22). *SATr: Slice Attention with Transformer for Universal Lesion Detection (MICCAI'22)

本项目为MICCAI22文章 "SATr: Slice Attention with Transformer for Universal Lesion Detection" 的开源代码，因为我们的方法具有通用性，我们仅开源SOAT的版本, A3D-SATr

Code structure

• main structure The A3D-SATr version is heavily based on the A3D work (MICCAI'21) and mmdetection The main structure please follow the above mentioned two repository.
• SATr structure Our code modifications mainly located in the Class Trans_with_A3D (line 753 in nn\models\truncated_densenet3d_a3d.py)

Installation

• git clone this repository
• pip install -e .

The code requires only common Python environments for machine learning. Basically, it was tested with Python 3 (>=3.6) PyTorch==1.3.1 numpy==1.18.5, pandas==0.25.3, scikit-learn==0.22.2, Pillow==8.0.1, fire, scikit-image Higher (or lower) pytorch versions should NOT work and the torch1.3.1 are not avaliable in the offical website. This is BUG from AD3 work. You can Use NEW URL channels,

1. Add the tsinghua URL channels to you conda permanently

conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/msys2/
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/conda-forge
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main/
conda config --set show_channel_urls yes

conda install pytorch=1.3.1 

2. Add the tsinghua URL channels to you conda temporary

conda install pytorch=1.3.1 -c https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/

How to run the experiments

• Dataset

  • Download DeepLesion dataset
  • Before training, mask should be generated from bounding box and recists. mask generation

• Preparing mmdetection script

  • Specify input ct slices in ./deeplesion/mconfigs/densenet_align.py through modifing NUM_SLICES in dict dataset_transform

  • Specify data root in ./deeplesion/ENVIRON.py

• A3D-SATr weights

  • Google Drive
  • 百度网盘 (w0lt)

*Suggest weight location:

/deeplesion/work_dirs/densenet_3d_trans_a3d_r2/

• Training

CUDA_VISIBLE_DEVICES=0 python -W ignore -m torch.distributed.launch  --nproc_per_node={num of GPUs} --master_addr 127.0.0.2 --master_port 2607 {train_dist files} {configs files} --launcher pytorch

e.g.

CUDA_VISIBLE_DEVICES=2 python -W ignore -m torch.distributed.launch  --nproc_per_node=1 --master_addr 127.0.0.2 --master_port 2607 deeplesion/train_dist.py deeplesion/mconfigs/densenet_a3d.py --launcher pytorch

• Test

CUDA_VISIBLE_DEVICES=0 python eval.py --config {configs files} --checkpoint {checkpoint files}

e.g.

CUDA_VISIBLE_DEVICES=0 python eval.py --config /deeplesion/mconfigs/densenet_a3d.py --checkpoint /deeplesion/work_dirs/densenet_3d_trans_a3d_r2/epoch_19.pth

Citation

bib:

@article{li2022satr,
title={SATr: Slice Attention with Transformer for Universal Lesion Detection},
author={Li, Han and Chen, Long and Han, Hu and Zhou, S Kevin},
journal={arXiv preprint arXiv:2203.07373},
year={2022}
}