Image-to-Graph Transformation via Superpixel Clustering to Build Nodes in Deep Learning for Graph

Abstract

In recent years, convolutional neural network (CNN) becomes the mainstream image processing techniques for numerous medical imaging tasks such as segmentation, classification and detection. Nonetheless, CNN is limited to processing of fixed size input and demonstrates low generalizability to unseen features. Graph deep learning adopts graph concept and properties to capture rich information from complex data structure. Graph can effectively analyze the pairwise relationship between the target entities. Implementation of graph deep learning in medical imaging requires the conversion of grid-like image structure into graph representation. To date, the conversion mechanism remains underexplored. In this work, image-to-graph conversion via clustering has been proposed. Locally group homogeneous pixels have been grouped into a superpixel, which can be identified as node. Simple linear iterative clustering (SLIC) emerged as the suitable clustering technique to build superpixels as nodes for subsequent graph deep learning computation. The method was validated on knee, call and membrane image datasets. SLIC has reported Rand score of 0.92±0.015 and Silhouette coefficient of 0.85±0.02 for cell dataset, 0.62±0.02 (Rand score) and 0.61±0.07 (Silhouette coefficient) for membrane dataset and 0.82±0.025 (Rand score) and 0.67±0.02 (Silhouette coefficient) for knee dataset. Future works will investigate the performance of superpixel with enforcing connectivity as the prerequisite to develop graph deep learning for medical image segmentation.