Applying Graph neural networks for medical image analysis

Project Motivation

GNNs can learn spectral and spatial features from the data, they have less trainable parameters and thus possibly are more explainable. Also, since it is hard to define localized convolutional filters for non-Euclidean data, GNNs can potentially perform better than existing Convolutional Neural Networks. Unlike CNNs, GNNs can work on variable sizes of graphs. 

Project Description

Graphs are abstract data structures comprising nodes and edges, where edges denote the relationship between nodes. Graphs have more expressive power as they can be used to model large systems like social networks, protein-protein interactions, and much more. Graph neural networks are models that try to harness the power of graph structure. For example, an image is not the right data to model a 3D object, a point cloud which is a collection of points that represents a 3D object is the perfect data for such modelling. Most of the known deep learning methods are unsuitable for such data types, whereas in such domains GNNs perform well because they can be applied to data in non-Euclidean space. Our objective is to apply GNNs to the medical imaging domain, to get a more robust and more explainable model for various modalities in medical imaging like MRI (Magnetic Resonance Imaging), CT (Computed Tomography), X-RAY, etc. 

Status: Ongoing

Team Members

Partners