This project is for constructing a simple 3d-based convolutional neural network for brain imaging data. The trained model would be used for developing a highly interpretable deep leanring method for 3d brain imaging data. Currently, this repo only focuses on building a predictive neural network.
The data were processed DTI imaging data (FA ,MD and other maps) and were not shared online. The dataset created by the package torchio directly fits torch.utils.data.DataLoader. Currently we have 1599 available subjects for training. The default setting for data transformation is None . Please refer to https://torchio.readthedocs.io/ for any infomation about transformation. (Pérez-García et al., TorchIO: a Python library for efficient loading, preprocessing, augmentation and patch-based sampling of medical images
in deep learning. Link: https://arxiv.org/abs/2003.04696).
Note: although we have a matched criterion for training set. We did not apply it in this version. A matched training set will be added soon.
Python 3.7 + Tensorflow Pytorch CUDA 10.1 CuDNN
start training:
python main.py train --load=False
set up the dashboard for visualization:
tensorboard --logdir=./logs/exp_mm-dd/ #mm-dd:month-day, e.g. 09-02
08/29/20 The model is still being training under different architectures.