DGA-Net: A Dual-branch Group Aggregation Network for Liver Tumor Segmentation in Medical Images

Hepatocellular carcinoma (HCC) is one of the most common malignant tumors worldwide. Accurate segmentation of the liver and lesion areas is crucial for diagnosis, surgical planning, and rehabilitation therapy. Although deep learning technologies have been applied to the automatic segmentation of the liver and tumors, there are still issues such as insufficient utilization of inter-pixel relationships, lack of refined processing after the fusion of high-level and low-level features, and high computational costs. To address the issues of lacking inter-pixel dependency and high model parameter requirements, we propose a Dual-branch Group Aggregation Network for Liver Tumor Segmentation in Medical Images (called DGA-Net). The network consists of a dual-branch encoder comprising the Fourier Spectral Learning Multi-Scale Fusion (called FSMF) branch and a Multi-Axis Aggregation Hadamard Attention (called MAHA) branch, as well as a decoder with a Group Multi-Head Cross-Attention Aggregation (called GMCA) module. The FSMF branch utilizes a Fourier network to learn amplitude and phase information, capturing richer features and details. The MAHA branch combines spatial information to enhance discriminative features while effectively reducing computational costs. The GMCA module improves localization capabilities and establishes long-range inter-pixel dependencies by merging features from different branches. Experiments on the public LiTS2017 liver tumor dataset show that the proposed method outperforms existing state-of-the-art approaches, achieving segmentation accuracies of 95.23% for the liver and 69.51% for tumors in DPC (Dice Per Case). Additionally, experiments on other datasets demonstrate that this method delivers excellent results, highlighting its strong generalization ability.