A novel deep learning-based approach for liver segmentation from 3D CT images using probabilistic shape guidance
TL;DR
In comparison to the current state-of-the-art deep learning-based segmentation models, the proposed system surpasses them, highlighting its potential and reliability in precisely segmenting liver regions in CT volumes.
Abstract
Accurate liver segmentation is essential for computer-aided diagnosis (CAD) systems, yet it remains challenging due to anatomical variability, indistinct organ boundaries, and tissue intensity variations in both normal and pathological cases. To address these challenges, a deep learning-based framework is proposed for segmenting liver regions from 3D computed tomography (CT) volumes. Initially, this system creates a 3D probabilistic shape map to estimate the probability of each voxel being a liver voxel, utilizing a reference atlas dataset. To facilitate this, the atlas references are first aligned with the input CT volume via 3D affine registration, establishing a coarse but essential spatial correspondence. Following this alignment, the model adaptively matches each voxel in the input CT volume with cubic neighborhoods of varying sizes derived from the aligned liver atlas to compute the final shape map probabilities. Finally, a dual-path UX-Net (DPUX-Net) architecture is introduced which receives the original CT volume, combined with its corresponding probabilistic shape map, as input, and produces accurate and anatomically consistent liver segmentations. The efficacy of the proposed liver segmentation system is assessed on 120 CT volumes using multiple quantitative metrics, including the Dice similarity coefficient (DSC), Jaccard index, absolute volume difference (AVD), and Hausdorff distance (HD), achieving scores of $$93.19\%\pm 2.53\%$$ , $$87.12\%\pm 4.42\%$$ , $$4.53\pm 3.20$$ , and $$7.37\pm 5.19$$ , respectively. The evaluation includes both hold-out test samples from the training dataset and entirely unseen external datasets, demonstrating the robustness and generalizability of the proposed framework. In comparison to the current state-of-the-art deep learning-based segmentation models, the proposed system surpasses them, highlighting its potential and reliability in precisely segmenting liver regions in CT volumes.