An image processing application for liver tumour segmentation

Abstract

Liver cancer is one of the leading causes of death worldwide. Consequently, the development of accurate and reliable segmentation techniques is indispensable for tumour volume measurement and staging analysis. An interactive algorithm for liver tumour segmentation was developed, allowing the user to quickly paint the object of interest in the image using an intelligent paintbrush. This technique was based on an image partitioning into homogeneous primitives regions by applying a pseudo-watershed algorithm on an image gradient magnitude. Outcome of this initial segmentation was the input of an efficient region merging process to find the best image partitioning, based on the minimum description length principle. The algorithm was evaluated on Computed Tomography (CT) and Magnetic Resonance (MR) data using the dice similarity coefficient (DSC) as a statistical validation metric. This led to a DCS mean scores of 87% and 84% on the CT and MR studies, respectively.