We set out to (1) connect the novel graph machine learning approaches with the existing basic generative adversarial network architecture. (2)Pursue un-trialed graph machine learning previous approaches for effectively capturing the topological features of the brain connectivity matrices. (3)Propose, implement and evaluate the superiority of the various novel combinations of the relevant elements. As the main component of the GAN, we implement a novel architecture that combines both superiority of GNNs and node spatiality. In that case, it not only improves the accuracy of the Alzheimer's Disease(AD) classification task but also proves the availability of adding graph neural network approaches to the existing GAN. To perform the quantitative analysis, our method was tested on three different Brain MRI datasets collected for binary dementia classification tasks.
All codes were written on Google Collaboratory. It should be convenient for everyone to run the code or observe the outputs.
In this project, we have utilized three different datasets, AD, ADNI, and CNI. All datasets were anonymised to avoid identifying personal information by inverse engineering. In other datasets folders, it contains the code of classification tasks with different methods we have tried. The node location folder represents the node coordinates of AD and ADNI datasets.
The graph neural network GAN is also written in the Google Collaboratory. It contains both training, testing, and visualisation parts of the project. The pkl file is the trained generator model.