nnHCCNeXt: A Lightweight Algorithm for Hepatocellular Carcinoma Segmentation in 3D Venous Phase CT Imaging

Hepatocellular carcinoma (HCC) is the third leading cause of cancer-related mortality globally, and its diagnosis and treatment rely heavily on precise CT imaging, with tumor segmentation being critical for treatment planning. Current segmentation methods struggle with the heterogeneous radiological presentation of HCC lesions in contrast-enhanced CT images. We developed nnHCCNeXt, an enhanced 3D U-Net architecture integrating two novel components: (1) HCCNeXt blocks employing multi-scale inception depthwise convolutions for feature extraction, and (2) Residual group attention gate (RGAG) blocks for improving feature flow. The model was trained on a dataset containing 338 venous phase CT scans with expert annotations and validated on a separate set of 49 cases. A comparative analysis was conducted, including architecture ablation studies and comparisons with six state-of-the-art models. Our approach achieved superior segmentation performance with Dice scores of 0.8101 and HD95 of 20.8102, representing 7.97% and 27.14% improvement over nnU-Net. The model maintains 5.52 million parameters, which is 91.2% fewer than SwinUNETR. Ablation studies confirmed the individual contributions of HCCNeXt blocks and RGAG modules. The nnHCCNeXt framework demonstrates a remarkable balance between segmentation precision and computational efficiency, outperforming existing models and providing a promising solution for automatic segmentation in HCC CT imaging.