Random Walks with GMM Statistical Inference Algorithm for Segmentation of Ground Glass Opacity Pulmonary Nodules

Abstract

Segmentation of Ground Glass Opacity (GGO) nodules is still an emerging and challenging research problem due to inhomogeneous interiors, blurred boundaries, and irregular shapes across different patients. In order to segment GGO pulmonary nodules, a novel random walk with Gaussian Mixture Model statistical inference (known as RW_GMM) algorithm is proposed in this paper. During the graph construction step, intensity, texture and spatial distance features are incorporated to calculate a new affinity matrix. The spatial distance between neighboring nodes as a weight is introduced to penalty the inconsistency of features. Spatial statistical information is encoded by building GMM models of foreground and background and the fuzzy membership value is calculated as a weight to penalize the inconsistence between the predefined probabilities and the calculated probabilities. The experimental results on the LIDC dataset show the superior performance of the proposed algorithm.