LiT-HiSegFormer-Net: A hybrid SegFormer network with Bi-level context fusion for liver and tumor segmentation in CT images

Precise segmentation of the liver and associated tumors is vital for diagnosing and treatment planning of hepatocellular carcinoma, particularly in radiation therapy. Although U-Net is a CNN-based architecture that has accomplished remarkable achievement in biomedical image segmentation, its dependence on convolutional operations limits its capacity to extract long-range contextual information and spatial dependencies in CT images with low tissue contrast and blurred organ boundaries. Transformer-based models address this issue by modeling global context but often fail to extract low-level features essential for precise localization. Since both global and local level features are crucial for accurate segmentation in complex anatomical scenarios, we propose LiT-HiSegFormer-Net, a U-shaped encoder–decoder architecture that integrates the advantages of CNNs and Transformers. A CNN and Swin Transformer are used in the encoder to extract required features, which are fused using a cross-attention-based fusion module in the skip connections. Experiments conducted on the LiTS-ISBI and 3D-IRCADb datasets indicate the efficacy of the presented method, achieving enhanced performance in liver organ and tumor segmentation compared to existing approaches. • Hybrid U-shaped architecture that integrates CNN-based local feature extraction with Swin Transformer-based global context modeling for liver and tumor segmentation. • An approach that fuses the global context of transformers with local features of CNNs using bidirectional cross-attention formulation. • Skip connection embedding of the bidirectional cross-attention module to strengthen the detection of small tumors.