Hybrid Attention-Transformer U-Net Model for Liver Tumor Segmentation

Liver Tumor is one of the main causes of millions of human deaths. Early detection and segmentation of these liver tumors using CT scans could prevent millions of patient’s deaths. Currently, Deep learning is highly proving its significance in medical image extractions. The U Net, one of the popular network models that modifies the connections between the nodes and adjusts the structure of convolution units. This can lead to some output features of convolution units being ignored. This paper proposes a Hybrid Attention Transformer U Net model which combines Attention mechanism and Transformer block within U Net. This model captures both local and global contextual information which improves the depiction of complex anatomical structures. The effectiveness of proposed hybrid model is evaluated using the public dataset: LITS. The Dice Similarity Coefficient achieved by the proposed model is 92.8% for liver tumor segmentation with a Volumetric Overlap error of 0.057 and Relative volumetric difference of 0.0029. The segmentation results demonstrate that the proposed hybrid network outperformed the previous neural networks.