Liver vessel segmentation based on attention guided deep convolutional neural network from CT images

Background and aims: Liver vessel segmentation from abdominal computed tomography (CT)images is crucial for liver cancer surgery and liver transplants, given the significant impact ofhepatic vascular structure.The task of liver vessel segmentation is of particular complexity, very challenging, and time-consumingbecause of the large anatomical variability in the size, position, and branching of theportal and hepatic veins. While conventional methods have been useful in liver vessel segmentation,deep learning-based methods have been shown to provide superior performance. Therefore,automatic liver vessel segmentation using deep learning based approaches became a fundamentalprocessing step of computer-aided diagnosis, which aims to increase precision in surgical planningand has become increasingly popular in recent years. These approaches can reduce manualinteractions and greatly simplify the work of physicians. Therefore, in this paper, a novel automaticmethod for liver vessel segmentation based on deep learning approaches is presented.Method: We propose an end-to-end vessel segmentation network including multi-head attentionby expanding a deep convolutional neural network to employ an effective combination of convolutionand self-attention.The network architecture incorporates multi-scale convolutional operators to capture local spatialinformation. Additionally, we designed the Attention-Guided Concatenation module to adaptivelyselect context features from low-level features based on guidance from high-level features. Theproposed method was thoroughly evaluated on the ۳Dircadb and the MICCAI ۲۰۱۸ Medical SegmentationDecathlon (MSD) Challenge datasets.Results: The experiment results of liver vessel segmentation on abdominal CT images demonstratethat the proposed method can effectively segment liver vessels with a dice score of ۷۴.۶۱%and outperforms previous methods by up to ۷%.Conclusion: The proposed network structure is highly effective in distinguishing between thevessel and non-vessel regions, resulting in accurate liver vessel segmentations.

کلیدواژه ها:

Deep Learning ، Liver Vessel Segmentation ، Attention Guided ، Convolutional Neural Network