This repository provides code for an efficient transfer learning method designed for structural MRI scans of brains.
As a deep learning model, we use the Simple Fully Connected Neural Networks (SFCNs) by Peng et al (2021) As pre-training data set, we use the UK Biobank neuroimaging data set giving T1-weighted structural MRI data. As target data sets, we use OASIS-3, IXI, and ABIDE I and II.
We also provide a tool to visualize the features of a neural network via t-SNE . On the right, you can find an example of a "brain manifold" created by the convolutional kernels of the network and visualized via t-SNE. Below, there are further examples.
To use this repository, one can simply clone it and install requirements specified in requirements.txt.
Please note this installs additional packages used for visualization. In particular, we use PyTorch
as a library for automatic differentation.
The main execution of one our methods is done via executing
main.py -deb full -con /path/to/json_file.json
where full can be changed to debug for debugging. Thejson_file.json can be created via input.py.
The core implementation of SFCNs can be found in /sfcn/. The folder /methods/ contains all methods which we consider: transfer learning methods, direct training, elastic net regression, and direct transfer of a pre-trained model.
I want to thank Stephen Smith (fMRIB) and Han Peng (VGG Group, fMRIB) for supervision of the project.
