Research on 3D liver blood vessel extraction technology based on deep learning

The three-dimensional liver vessel extraction is a medical image segmentation technique aimed at automatically extracting vascular structures within the liver from CT or MRI scans[1], providing reliable support for pre-surgical planning. This paper primarily introduces a liver vessel extraction method implemented using the PyTorch framework. Based on the 2D U-Net network, and referencing the 3D U-Net proposed by Çiçek Ö et al.[2], a 3D U-Net network is designed and constructed for this experiment. To address the under-segmentation issue of small vessels in 3D U-Net, residual blocks (Res Block) are introduced in both the encoder and decoder stages, forming the 3D Res-UNet network. Experimental results demonstrate that, with the same number of training iterations and hyperparameter settings, the 3D Res-UNet network achieves a 2.6% improvement in the Dice coefficient compared to the 3D U-Net. To broaden the receptive field and capture contextual information at multiple scales, a 3D Spatial Pyramid Pooling (3D ASPP) module with multi-scale dilation rates is incorporated into the 3D Res-UNet network.