3D Binary Lesion Mask Parsing

Abstract

Liver lesion segmentation is a key module for an automated liver disease diagnosis system. Numerous methods have been developed recently to produce accurate 3D binary lesion masks for CT scans. From the clinical perspective, it is thus important to be able to correctly parse these masks into separate lesion instances in order to enable downstream applications such as lesion tracking and characterization. For the lack of a better alternative, 3D connected component analysis is often used for this task, though it does not always work, especially in the presence of confluent lesions. In this paper, we propose a new method for parsing 3D binary lesion masks and an approach to evaluating its performance. We show that our method outperforms 3D connected component analysis on a large collection of annotated portal-venous phase studies.