HI-Net: Liver vessel segmentation with hierarchical inter-scale multi-scale feature fusion

Automated and accurate liver vessel segmentation is essential for clinical diagnosis and treatment, but the segmentation of small vessels still remains challenging due to their intricate topology and image noise. Most existing methods focus on reducing information loss during multiple single down-sampling steps, often neglecting to utilize multi-scale contextual information. This restricts the ability of the decoding process to capture contextual information from multiple receptive fields, resulting in the loss of high-level semantic details. To address this issue, in this paper, we propose a hierarchical inter-scale multi-scale feature fusion network for liver vessel segmentation called HI-Net. It includes a hierarchical multi-scale feature fusion module and several inter-scale dense connections to integrate different levels of feature information and mitigate the potential loss of high-level semantic information. In addition, deep supervision is also introduced to accelerate network convergence and enhance its ability to learn semantic features. Extensive experiments were conducted on the publicly available 3Dircadb dataset for liver vessel segmentation. The results demonstrated remarkable performance with 75.36% dice and 78.95% sensitivity, surpassing existing advanced liver vessel segmentation methods.