Abstract
The highly heterogeneous and irregular morphology of liver tumors presents considerable challenges for automated segmentation. To better capture complex tumor structures, this study proposes a liver tumor segmentation framework based on multi-scale deformable feature fusion and global context modeling. The method incorporates three key innovations: (1) a Deformable Large Kernel Attention (D-LKA) mechanism in the encoder to enhance adaptability to irregular tumor features, combining a large receptive field with deformable sensitivity to precisely extract tumor boundaries; (2) a Context Extraction (CE) module in the bottleneck layer to strengthen global semantic modeling and compensate for limited capacity in capturing contextual dependencies; and (3) a Dual Cross Attention (DCA) mechanism to replace traditional skip connections, enabling deep cross-scale and cross-semantic feature fusion, thereby improving feature consistency and expressiveness during decoding. The proposed framework was trained and validated on a combined LiTS and MSD Task08 dataset and further evaluated on the independent 3D-IRCADb01 dataset. Experimental results show that it surpasses several state-of-the-art segmentation models in Intersection over Union (IoU) and other metrics, achieving superior segmentation accuracy and generalization performance. Feature visualizations at both encoding and decoding stages provide intuitive insights into the model's internal processing of tumor recognition and boundary delineation, enhancing interpretability and clinical reliability. Overall, this approach presents a novel and practical solution for robust liver tumor segmentation, demonstrating strong potential for clinical application and real-world deployment.