An Improved Hybrid Algorithm for Accurate Determination of Parameters of Lung Nodules with Dirichlet Boundaries in CT Images

Background/Objectives: Segmentation of lung nodules with irregular boundaries in CT images remains a challenge. This work aims to advance the accuracy of segmentation using Random walker and watershed algorithm. Methods/Statistical Analysis: The input image is considered as a graph representing each pixel as a node. Two seed points which are user-defined (pre-labeled) pixels given as labels, one for the foreground and the other for the background. The gradient of the seed points are calculated. Then the probability of reaching the labeled pixels from each of the unlabeled pixels is obtained and a vector of probabilities is defined for each of the unlabeled pixels. Findings: The calculated vector of probabilities for each unlabeled pixel is combined and they can be assigned to one of the labels using the watershed algorithm to obtain tumor segmentation. We used 23 images for validating our method and our experiment compared the original random walker algorithm, random walker with improved weights and watershed segmentation results. Resulted images have maximum of DSC values as 0.92 for Random Walk, 0.94 for Random Walk with Improved Weight and 0.97 for Watershed combined. Applications/Improvements: Accurate segmentation of nodules with dirichlet boundaries in CT images using minimum number of seed points.