Multi-Scale Convolutional Attention and Structural Re-Parameterized Residual-Based 3D U-Net for Liver and Liver Tumor Segmentation from CT

Abstract

Accurate segmentation of the liver and liver tumors is crucial for clinical diagnosis and treatment. However, the task poses significant challenges due to the complex morphology of tumors, indistinct features of small targets, and the similarity in grayscale values between the liver and surrounding organs. To address these issues, this paper proposes an enhanced 3D UNet architecture, named ELANRes-MSCA-UNet. By incorporating a structural re-parameterized residual module (ELANRes) and a multi-scale convolutional attention module (MSCA), the network significantly improves feature extraction and boundary optimization, particularly excelling in segmenting small targets. Additionally, a two-stage strategy is employed, where the liver region is segmented first, followed by the fine-grained segmentation of tumors, effectively reducing false positive rates. Experiments conducted on the LiTS2017 dataset demonstrate that the ELANRes-MSCA-UNet achieved Dice scores of 97.2% and 72.9% for liver and tumor segmentation tasks, respectively, significantly outperforming other state-of-the-art methods. These results validate the accuracy and robustness of the proposed method in medical image segmentation and highlight its potential for clinical applications.