Precise Segmentation of Pathological Liver in 3D CT Scans Using Adaptive Shape Guidance from Prior Appearance

Abstract

This paper introduces a novel approach for extracting pathological liver regions from 3D computed tomography (CT) images. The proposed method leverages deep learning techniques to precisely segment pathological livers from 3D CT images. This is particularly challenging due to the inhomogeneity and diffused boundaries in the appearance of pathological liver tissue, as well as anatomical complexities such as shape and size variations and the liver’s close proximity to other organs. The proposed approach begins with 3D affine-based registration to roughly maximize the overlap between the subject to be segmented and the created liver atlas. Next, a 3D probabilistic shape map of the liver is generated by adaptively matching the appearance of each voxel with the prior liver atlas. This is achieved by selecting cubic neighbors of varying sizes around the location of each voxel. Finally, the original CT data, along with their corresponding 3D shape maps, are input into the 3D UX-Net. The proposed system is evaluated on 40 patients (20 with normal livers and 20 with pathological livers) using various metrics, including Dice similarity coefficient, overlap coefficient, absolute volume difference, and Hausdorff distance. It achieved scores of 94.66±1.1%, 89.88 ± 1.96%, 3.26± 2.66, and 7.18 ±2.65, respectively, outperforming different deep learning-based segmentation systems.