Segmentation of nodular medulloblastoma using random walker and hierarchical normalized cuts

Abstract

Medulloblastoma (MB) is the most common brain tumor in children. Recent studies have demonstrated a relationship between specific signaling pathway abnormalities, a tendency to more favorable outcomes, and a histopathological feature: nodular growth patterns. In this work we present a new segmentation scheme which requires minimal user interaction to segment nodules on MB histopathological sections. Our segmentation scheme consists of two steps: (1) color reduction using Hierarchical Normalized Cuts (HNCut), (2) Random Walker (RW) segmentation within the reduced HNCut color space. Across a cohort of 18 nodular MB images, our integrated HNCut and RW scheme yielded nodule segmentations with a Dice coefficient of 83:55 ± 12:4% and Predictive Positive Value (PPV) of 93:71 ± 9:0%. © 2011 IEEE.