Multi-Scale Residual Gated Attention U-Net for Liver Tumor Segmentation.

Accurate early diagnosis and treatment planning of liver cancer relies on precise liver and tumor segmentation in medical images. However, liver tumor segmentation faces complex anatomical structures, large lesion scale differences, and fuzzy lesion boundaries. Vanilla U-Net has three inherent structural drawbacks: inadequate cross-scale feature interaction in the encoder, semantic mismatch in skip connections, and insufficient multi-scale feature aggregation in the decoder. Therefore, this paper proposes a Multi-scale Residual Gated Attention U-Net (MRGA-UNet) for liver tumor segmentation. The Multi-Residual Coordinate Enhancement (MRCE), Cascaded Residual Gating (CRG) and Multi-Scale Feature Fusion (MFF) form an end-to-end feature pipeline tailored for liver CT segmentation. Coordinated feature propagation across the three modules adapts to the imaging traits of liver CT. This systematic matching across modules forms the overall novelty relative to mainstream residual and attention-based U-Net variants. The MRCE module is embedded in the encoder and combines multi-residual branches with coordinate attention to extract abundant multi-scale tumor features, tackling insufficient cross-scale interaction. The CRG module is arranged on skip connections; guided by high-level semantics, it optimizes features bottom-up to eliminate semantic mismatch between layers. The MFF module is deployed in the decoder to achieve sufficient multi-scale aggregation and reinforce tumor feature representation. Tested on the LiTS and 3DIRCADb datasets, the proposed model achieves a Dice Similarity Coefficient (DSC) of 83.3% and 81.8%, respectively. Compared with MedNeXt, the proposed model increases DSC by 5.3 percentage points and reduces VOE by 0.069 on the LiTS dataset. On the 3DIRCADb dataset, the proposed method achieves a 7.0 percentage points improvement in DSC and reduces VOE by 0.080. The proposed model achieves superior segmentation performance primarily in DSC and VOE, while maintaining competitive performance on other metrics. Ablation studies confirm that each proposed module contributes positively to the overall segmentation performance.