Abstract
Liver and tumor segmentation is an important technology for the diagnosis of hepatocellular carcinoma. However, most existing methods struggle to accurately delineate the boundaries of the liver and tumor due to significant differences in their shapes, sizes, and distributions, which leads to unclear segmentation of the liver contour and incorrect delineation of the lesion area. To address this gap, we propose a hybrid gabor attention convolution and transformer interaction network with hierarchical monitoring mechanism for liver and tumor segmentation, named HyborNet. Generally, the proposed HyborNet consists of a local and a global feature extraction branch. Specifically, the local feature extraction branch consists of several cascaded gabor attention convolutional blocks, each of which contains a multi-dimensional interactive attention module and a gabor convolutional module. In this way, fine-grained information about the liver and tumor can be extracted, which refines the edge details of the target area and accurately depicts the lesion area. The global feature extraction branch is constructed with a transformer model, which is capable of extracting coarse-grained information about the liver and tumor and accurately distinguishing them from similar tissues. Additionally, we propose a cross-attention-based dual-branch interaction module that adaptively fuses features from different perspectives to emphasize the target region, thereby enhancing the network's segmentation performance. Finally, a hierarchical monitoring mechanism is employed in the decoding stage, which provides additional feedback from deeper intermediate layers to optimize the segmentation results. Extensive experimental results demonstrate that HyborNet significantly outperforms other state-of-the-art models in liver and tumor segmentation tasks. The proposed model effectively enhances liver image segmentation accuracy, assisting doctors in making more precise diagnoses.