basiralab/tis-net

Intermodalitty graph superresolution.

Template-Based Inter-modality Super-resolution of Brain Connectivity (TIS-Net)

TIS-Net is a framework for super-resolving cross-modality brain graphs using one shot learning (i.e., a single training sample).

Please contact [email protected] for inquiries. Thanks.

Introduction

This work is accepted at the PRIME-MICCAI 2021 workshop.

Template-Based Inter-modality Super-resolution of Brain Connectivity
Furkan Pala^1,†, Islem Mhiri^1,2,†, and Islem Rekik^1,*
1 BASIRA Lab, Faculty of Computer and Informatics, Istanbul Technical University, Istanbul, Turkey
² Laboratory of Advanced Technology and Intelligent Systems, National Engineering School of Sousse, Tunisia
^* corresponding author: [email protected], https://basira-lab.com/
^† co-first authors

Abstract: Brain graph synthesis becomes a challenging task when generating brain graphs across different modalities. Although promising, existing multimodal brain graph synthesis frameworks based on deep learning have several limitations. First, they mainly focus on predicting intra-modality graphs, overlooking the rich multimodal representations of brain connectivity (inter-modality). Second, while few techniques work on super-resolving low-resolution brain graphs within a single modality (i.e., intra), inter-modality graph super-resolution remains unexplored though this avoids the need for costly data collection and processing. More importantly, all these works need large amounts of training data which is not always feasible due to the scarce neuroimaging datasets especially for low resource clinical facilities and rare diseases. To fill these gaps, we propose an inter-modality super-resolution brain graph synthesis (TIS-Net) framework trained on one population driven atlas namely-connectional brain template (CBT). Our TIS-Net is grounded in three main contributions: (i) predicting a target representative template from a source one based on a novel graph generative adversarial network, (ii) generating high-resolution representative brain graph without resorting to the time consuming and expensive MRI processing steps, and (iii) training our framework using one shot learning where we estimate a representative and well centered CBT (i.e., one shot) with shared common traits across subjects. Moreover, we design a new Target Preserving loss function to guide the generator in learning the structure of target the CBT more accurately. Our comprehensive experiments on predicting a target CBT from a source CBT using our method showed the outperformance of TIS-Net in comparison with its variants. TIS-Net presents the first work for graph synthesis based on one representative shot across varying modalities and resolutions, which handles graph size, and structure variations. Our Python TIS-Net code is available on BASIRA GitHub at https://github.com/basiralab/TIS-Net

Code

This project was implemented using Python 3.6.9 on a Linux platform.

Installation

Create a virtual env using the following command

python3 -m venv TIS_NET

Activate the virtual env using the following command

source TIS_NET/bin/activate

Use pip to install dependencies. Note that we developed the project on a CUDA supported system. Our CUDA version is 10.1, so the installion commands below include cu101.

pip install torch==1.8.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html
pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.8.0+cu101.html
pip install torch-sparse -f https://pytorch-geometric.com/whl/torch-1.8.0+cu101.html
pip install torch-geometric
pip install matplotlib
pip install fbpca
pip install annoy

For CPU installation or another version of CUDA or PyTorch, check out PyTorch Geometric Docs and PyTorch Website.
Also our framework depends on SIMLR framework of which source code is also included in this repository in /SIMLR_PY folder.

Running TIS-Net

To run TIS-Net, use the following command

python3 demo.py

Make sure that your virtual env is activated.

You can edit the config.py file according to your needs by changing hyperparameters such as number of epochs for training, and you can feed an external dataset or change the shape of the simulated data.

Data format

TIS-Net framework needs source and target data to run. Each of these data represents a bunch of subjects and their brain graphs (networks). Note that our framework is based on single-view brain graphs. A brain graph can be represented as an adjacency matrix with shape

[num_ROIs x num_ROIs]

where num_ROIs is number of region of interests in the brain graph, different for source and target datasets since our framework offers an inter-modality mapping. However, a more compact way to store this matrix is vectorizing its upper-triangular part so that this matrix can be represented as a vector with

num_ROIs x (num_ROIs - 1) / 2

elements. Let us define this number as num_features since, intuitively, each subject has this amount of features.

Now, we can specify the shape of source and target datasets which are

[num_subjects x num_features]

num_features is different for source and target datasets.

You can either supply an external dataset by assigning DATASET=E in config.py file or use a simulated dataset (created using random normal distribution) by specifying number of ROIs in both source and target domains via changing N_SOURCE_NODES and N_TARGET_NODES variables, respectively. You can also specify number of subjects with N_SUBJECTS variable.

Components of DGN’s Code

Component	Content
config.py	Includes hyperparameter and other options. You may modify it according to your needs.
model.py	Implementation of the model.
demo.py	Driver code that import variables from config.py, trains and tests TIS-Net using k-fold cross-validation.
helper.py	Includes some helper functions such as preprocessing data and printing functions.
centrality.py	Includes functions for calculating pagerank, eigenvector and betweenness centralities. These are called topological measures in our project.
/models 1	Stores the weights of trained models under the /MODEL_NAME folder. MODEL_NAME can be specified in config.py.
/output 1	Stores the figures of predicted and ground-truth CBTs together with residuals under the /MODEL_NAME folder.
/data 1	Stores the generated simulated data under the /MODEL_NAME folder.

¹ This is an auto-generated folder and created when the program is executed.

Example Result

The figure above shows an example result of predicted and ground-truth CBT together with residual between them trained and tested using a simulated data, source domain with 35 ROIs and target domain with 160 ROIs.

• MAE: Mean Absolute Error
• PR: Pagerank centrality
• EC: Eigenvector centrality
• BC: Betweenness centrality

¹ Calculated between predicted and ground-truth CBT.

² Calculated between predicted CBT and each test sample in target domain.

YouTube videos to install and run the code and understand how TIS-Net works

To install and run TIS-Net, check the following YouTube video:

https://www.youtube.com/watch?v=c3HrY53JUw0

To learn about how TIS-Net works, check the following YouTube video:

https://www.youtube.com/watch?v=ydQvqx8rMiw

Please cite the following paper when using DGN

  @inproceedings{pala2021,
    title={Template-Based Inter-modality Super-resolution of Brain Connectivity (TIS-Net)},
    author={Pala, Furkan and Mhiri, Islem and Rekik, Islem},
    booktitle={International Workshop on PRedictive Intelligence In MEdicine},
    year={2021},
    organization={Springer}
  }