Combined Radiology and Pathology Classification

MICCAI 2020 Combined Radiology and Pathology Classification Challenge (1st place solution)

Hardware

• 4*NVIDIA Tesla P40 GPU cards
• 32GB of RAM

Pre-requisites:

• torch >=1.3.0, nibabel, batchgenerators, efficientnet_pytorch

Usage

Preparation

1. Data Preparation

• Download challenge data

2. Training Splits (spilt/train1.txt)

MRI training

1. 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

WSI training

This is similar to the previous MRI training

1. Trainning Glioblastoma/None Glioblastoma

cd pathology
python train_g.py

2. Trainning Oligodendroglioma/Lower grade astrocytoma

cd pathology
python train_ao.py

Reference

This is code based on MedicalNet

Citation

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}
}