Liver Blood Vessel Segmentation Research Based On Multi-scale Feature Analysis

Accurate segmentation of liver vessels is the base to build tools of surgical oncology planning and to realize medical visualization applications.The morphological structure and distribution of liver blood vessels have a significant guidance role for liver surgical treatment.Due to the low contrast of liver vessels with their surrounding tissues on CT images and the complex structure of liver vessels,accurate implementation of liver vessel segmentation is challenging,especially for small vessels.Artificial intelligence is booming,and deep learning methods highlight the powerful advantages and intelligence in the field of image segmentation.The 3D U-Net network model based on 3D convolutional U-shaped structure can make full use of the characteristics of 3D CT images and increase the precision in the segmentation of medical images.On the basis of this,combining with the characteristics of liver vessels,the thesis develops a segmentation network model research that conforms to liver vessels structure.The specific research contents are as follows.1.To address the problem that the fusion of coarse-grained and fine-grained multi-scale semantic information would be neglected in the network structure commonly used for segmentation and cannot effectively learn the multi-scale information of liver vessels,proposed a segmentation network model HPM-Net with progressive idea.This model mainly adopts a multi-stage structure,using 3D U-Net optimization network of different depths to obtain different size of the receptive fields,and enhances the representation of segmented vessels features by integrating coarse-grained location information and fine-grained semantic information of liver vessels into multi-scale predictions through the combined intra-stage internal progressive and inter-stage external progressive.The experimental results show that HPM-Net can achieve an average dice value of 75.18% on the public dataset 3Dircadb,which has higher accuracy and validity.2.A hierarchical inter-scale feature fusion network model HI-Net is proposed in order to be able to obtain vessel multi-scale features and to reduce the high-level semantic information lost in the process of recovering low-resolution features to original resolution.This model introduces a hierarchical multi-scale feature fusion module between the encoder and decoder,which can effectively combine the feature information of liver vessels at different scales.A dense inter-scale connection is used in the decoder part,which enables the model to reduce the loss of high-level semantic information of liver vessels during the up-sampling process.The experimental results show that the HI-Net network model reaches an average Dice value of75.36% on the public dataset 3Dircadb that is able to accurately segment the more complete liver vessels.