Hybrid Topology of Graph Convolution and Autoencoder Deep Network For Multiple Sclerosis Lesion Segmentation

The notion of deep learning is being revolutionised, e.g., the advent of graph convolution network (GCN) in the application of medical image classification. Though GCN is gaining significant popularity, the potential of GCN is not exploited to its maximum. This work proposes a new hybrid network based on convolution neural network (CNN) autoencoder and GCN. The proposed method has been applied to graph datasets which are designed by considering 3D medical resonance image (MRI) voxel as a node. CNN autoencoder is used to extract the imaging grid information and GCN learns these features in graph connectivity space. This hybrid network has been applied for the application of segmentation of lesions in multiple sclerosis (MS) diseases. This framework has been used for training on 30 MS patients of white matter lesions captured at the university medical center Ljubljana (UMCL) and validated on 20 longitudinal MS patient dataset of UMCL at one time instant. This novel framework improves the performance of dice similarity coefficient 85.5 % score in segmentation for three neighbors in graph data.