| The task of automatic segmentation of liver vessels has always been a topic of great concern in the field of medical image because of the significance of vessel segmentation for the diagnosis of disease and surgical planning.This thesis focuses on the automatic segmentation of liver vessels from the liver region in abdominal enhanced CT images.Due to low contrast,high noise,unstable image quality and complex shape of liver vessels,the segmentation performance of existing models is difficult to meet the actual needs.However,there are few publicly available liver datasets and the quality of the vessel labels is rough,which hinders the research of liver vessel segmentation algorithm.The existing vessel segmentation methods are mainly based on traditional image feature extraction and deep learning.The method based on image feature extraction extracted features from liver vessel images to obtain vessel pixel-level information.Different machine learning algorithms were used to distinguish blood vessels from non-blood vessels.However,in order to match the complex characteristics of liver vessels and improve the segmentation accuracy,multiple effective feature extraction methods need to be combined.Deep learning segmentation method obtains nonlinear connections between vascular features through a large amount of data training network,and independently learns to segment vessels.However,there are problems such as discontinuity of the segmented vessels,holes in the vessels,and small scale vessel leakage segmentation.Considering that doctors deal with medical images more often use the continuity of three-dimensional data to analyze the pathological information of two-dimensional images more intuitively and accurately,which requires higher data quality.Therefore,deep learning still has great research space in automatic vessel segmentation.In order to solve the above problems,this thesis improved the automatic segmentation algorithm of liver vessels in two directions of traditional machine learning and deep learning respectively.Firstly,a conventional segmentation algorithm based on image feature extraction is proposed.This algorithm is reduced dependence on a large number of vessel labels and the segmentation results are obtained by the XGBoost classifier identifying the vessel pixel points and then refining the segmentation results with Markov random fields,where the features provided to the classifier are obtained from the 22 filters used for image processing.Experiments conducted on the public dataset 3D-IRCADb show that the proposed model produces an average Dice score of 0.534,which is obviously superior than the results obtained from traditional machine learning methods.When this method was used on its own newly constructed dataset,Dice is significantly improved to 0.688,indicating the superiority of the newly constructed dataset and the effectiveness of this model in liver vessel recognition under the circumstance of limited data.Secondly,the common downsampling structure in existing deep learning models has some side effects on vessel segmentation,especially for small vessels.A bidirectional scaling module using only one downsampling followed by multiple upsampling is proposed in this thesis,where the downsampling amplifies the receptive field to obtain high-level features,the upsampling module amplifies the vessel details,and an attention module is added to the model to capture the natural connectivity of blood vessels.Experimental results show that the proposed method has made great progress in vessel segmentation,the dice obtained on the dataset 3D-IRcadb is 0.541,and the dice obtained on the newly constructed dataset is 0.704.Clearly,the newly constructed dataset and bidirectional scaling scheme are expected to be used to construct more accurate vessel segmentation models. |
