DS-UNet: enhancing liver tumor segmentation in CT images via dense state space selection and gating mechanism

Abstract

Early diagnosis and precise localization of malignant liver tumors are crucial for effective clinical decision-making. However, existing automated liver tumor segmentation methods for CT images still face the following challenges: (1) Traditional U-Net and its variants struggle to achieve accurate tumor localization and fail to resolve blurred segmentation boundaries in complex anatomical backgrounds; (2) CNN-based methods are limited by fixed local receptive fields, failing to model long-range contextual dependencies for small and morphologically heterogeneous tumors and (3) Existing mainstream methods fail to balance segmentation accuracy and computational efficiency in resource-limited clinical scenarios. To overcome them, we propose Dense State Space Selection U-Net, a dense state space selection network, to enhance liver tumor segmentation from CT images. By integrating a gating mechanism and dense state space blocks, DS-UNet effectively models spatial correlations and improves feature extraction, resulting in superior segmentation accuracy. Quantitative experiments on public datasets demonstrate the effectiveness, which achieves Dice coefficients of 74.76% and 74.04% on LITS and HCC datasets respectively. Besides, ours provides a feasible and accurate solution for automated liver tumor segmentation, contributing to advancements in medical image analysis. The code for this paper has been released at https://github.com/Fan-XYin/DS-UNet .