This is the code assoiciated with MICCAI 2020 paper titled "NAS-SCAM: Neural Architecture Search-Based Spatial and Channel Joint Attention Module for Nuclei Semantic Segmentation and Classification". In the code, we presents attention module architecture searching for nuclei semantic segmentation and classification. Which has two features: 1) the architecture of attention module can be automated search by NAS, 2) different architectures of attention modules can be searched in the same network. The dataset is from the challenge of MoNuSAC 2020. We make the preprocessed dataset available in Google Drive and Baidu Netdisk (extract code: lj4p). Simply changing the dir in train_functions.py can make the program properly run. We make the following preprocessing: 1) color normalization, 2) overlap cropping. A detailed description of our dataset will be provided soon.
- pytorch
- numpy
- optparse
- pillow
- nibabel
- tqdm
- scipy
Firstly, download dataset provided. Then, changing the default dataset path in function "get_args" in train.py, or adding parameter in comment line like below
python train.py -d '/home/username/final_dataset (add your path here)'
If you think this code can help you, please kindly cite our paper.
@inproceedings{liu2020scam,
title={NAS-SCAM: Neural Architecture Search-Based Spatial and Channel Joint Attention Module for Nuclei Semantic Segmentation and Classification},
author={Liu, Zuhao and Wang, Huan and Zhang, Shaoting and Wang, Guotai and Qi, Jin},
booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
pages={263--272},
year={2020},
organization={Springer}
}