Graph neural network (GNN) is used for unstructured data. I created a PPT for the general introduction on GNN.
Blogs to introduce GNN in general
GNN on medical images
We could regard each pixel (or voxel for 3D images) as a node in a graph, and then convert a image data to a graph data. Such conversion is memory efficient to save data and train networks for sparse data (like vessels).
Python packages for GNN
torch_geometric is a library built upon PyTorch to easily write and train Graph Neural Networks (GNNs). All the related documentation including public datasets, dataloaders, convolution layers could be found here and here.
I strongly recommend to learn the package from the official tutorials
Different GNN designs and their implementations could be found here
some tips
- The number of graph convolution layers is normally between 2 and 4. Deeper networks do not lead to better performance.
- Data augmentation is time-consuming for big graph data.
- GNN is super faster than CNN (about 10 times faster).
- You could read some surveys or overview papers to have a complete view about GNN.