MICCAI 2020 Combined Radiology and Pathology Classification Challenge (1st place solution)
- 4*NVIDIA Tesla P40 GPU cards
- 32GB of RAM
- torch >=1.3.0, nibabel, batchgenerators, efficientnet_pytorch
- Data Preparation
- Download challenge data
- Training Splits (spilt/train1.txt)
- Trainning Glioblastoma/None Glioblastoma(pretrain 3d medical Weights )
cd mri
python train_g.py
RESNET=False #False is resnet, True is densenet
model.conv1 = nn.Conv3d(4,....) #The input channel is 5 if the tumor segmentation region exists, otherwise it is 4
model = densenet.densenet121(first=5，..) #The input channel is 5 if the tumor segmentation region exists, otherwise it is 4
#datasets.brain.py
BrainDataset_AO,BrainDataset_G # AO dataset，G_dataset
return img_array[:4], labels #The input channel is 5 if the tumor segmentation region exists, otherwise it is 4
2.Trainning Oligodendroglioma/Lower grade astrocytoma
After the first stage of training, the second stage of training needs to use the weights trained in the first stage to warm up
cd mri
python train_ao.py
This is similar to the previous MRI training
- Trainning Glioblastoma/None Glioblastoma
cd pathology
python train_g.py
- Trainning Oligodendroglioma/Lower grade astrocytoma
cd pathology
python train_ao.py
This is code based on MedicalNet
Please use below to cite this paper if you find our work useful in your research.
title={Combining Radiology and Pathology for Automatic Glioma Classification},
author={Wang, Xiyue and Wang, Ruijie and Yang, Sen and Zhang, Jun and Wang, Minghui and Zhong, Dexing and Zhang, Jing and Han, Xiao},
journal={Frontiers in Bioengineering and Biotechnology},
volume={10},
year={2022},
publisher={Frontiers Media SA}
}