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. Due to its high invasiveness and poor prognosis, it ranks among the top three causes of cancer-related deaths globally. Accurate segmentation of the liver and lesion areas is crucial. It provides key support for diagnosis, surgical planning, and rehabilitation therapy. Deep learning technologies have been applied to the automatic segmentation of the liver and tumors. However, several issues remain, such as insufficient utilization of inter-pixel relationships, lack of refined processing after fusing high-level and low-level features, and high computational costs. To address insufficient inter-pixel modeling and high parameter costs, we propose DGA-Net (Dual-branch Group Aggregation Network for Liver Tumor Segmentation in Medical Images), a dual-branch architecture that includes two main components, i.e., a dual-branch encoder and a decoder with a specific module. The dual-branch encoder consists of the Fourier Spectral Learning Multi-Scale Fusion (FSMF) branch and the Multi-Axis Aggregation Hadamard Attention (MAHA) branch. The decoder is equipped with a Group Multi-Head Cross-Attention Aggregation (GMCA) module. The FSMF branch uses a Fourier network to learn amplitude and phase information. This helps capture richer features and details. The MAHA branch combines spatial information to enhance discriminative features. At the same time, it effectively reduces computational costs. The GMCA module merges features from different branches. This not only improves localization capabilities but also establishes long-range inter-pixel dependencies. We conducted experiments on the public LiTS2017 liver tumor dataset. Experiments on the public LiTS2017 liver tumor dataset show that the proposed method outperforms existing state-of-the-art approaches, achieving Dice-per-case (DPC) scores of 94.84% for liver and 69.51% for tumors, outperforming competing methods such as PVTFormer by 0.72% (liver) and 1.68% (tumor), and AGCAF-Net by 0.97% (liver) and 2.59% (tumor). We also carried out experiments on the 3DIRCADb dataset. The method still delivers excellent results, which highlights its strong generalization ability.