| The analysis of cerebrovascular angiographic image is of great significance for understanding the pathogenesis,preoperative diagnosis and treatment of various cerebrovascular diseases.In this paper,combining statistical models and deep learning methods,a deep study for cerebral vascular segmentation,cerebrovascular lesions localization,vascular centerline extraction,and preoperative interventional route planning is presented based on TOF-MRA(Time of flight-Magnetic Resonance Angiography)data.The following problems in this field:(a)although the statistical model has achieved sufficient development in cerebrovascular segmentation,due to the difference between imaging equipment and imaging parameters,there are significant differences between the data,which affects the robustness and segmentation performance of the statistical model.(b)deep learning has made significant progress in medical image processing,but it always faces the bottleneck of insufficient cerebrovascular labels in cerebrovascular segmentation.(c)During the preoperative interventional surgery,the clinicians need to repeatedly review the TOF-MRA image to determine the lesion location and interventional route.This process depends entirely on the doctor’s medical experience,which increases the doctor’s burden of diagnosing medical image and is not conducive to the accurate and rapid implementation of interventional treatment.Aiming the problems mentioned above,the main contributions and innovations of this paper are as follows:1.A device-independent FMM-MRF(Finite Mixture Model & Markov Random Field)cerebrovascular segmentation method is proposed.Skull stripping and histogram specification are used to reduce the region of interest and regulate the intensity distribution;Due to the small intracranial volume of vessels,the traditional EM algorithm is biased in the process of FMM parameter estimation.Therefore,an EM algorithm based on cerebrovascular knowledge is proposed;Probabilistic vascular feature maps is obtained by combining multi-scale vascular enhancement results with estimation vascular weights in FMM,then a dual-energy-constrained MRF process was constructed.2.A model-and data-driven method for cerebrovascular segmentation is presented.The device-independent FMM-MRF method is used to generate lots of cerebrovascular labels;then,the generality of cerebrovascular data is mined by deep learning,and a cleaning mechanism model is established to correct the wrong labels,so as to obtain high-quality labels;finally,according to the multi-scale characteristics of vessels,a convolutional neural network with two branches is constructed,which fully combines the advantages of dense connection and dilated convolution.3.Aiming at the location and segmentation of lesion targets,a two-stage convolution network structure is designed: A weighted cross-plane dilated convolution structure is formulated to replace the 3D convolution,which makes use of 3D context information,as well as reduce computation;meanwhile,under the same image resolution space,3D pyramid features are captured using densely connection and different dilated rate setting.The experiment shows that the method is better than Vnet.4.Aiming at the centerline extraction and intervention path planning,a new method for extracting the cerebrovascular centerline is proposed,which combines the thinning algorithm with the local edge distance map and multi-scale enhancement.At the same time,the centerline was transformed into a static vascular road network for the topological path analysis,and the preoperative intervention route is planned by Astar algorithm.The above mentioned contributions and innovations are of great significance to the current research on artificial intelligence methods for angiographic images analysis: further improving the robustness of statistical models and alleviating the highly dependence of deep learning on ground truths.For interventional embolization surgery,this work supports three-dimensional visualization of cerebral vessels and lesion location,thereby improving the personalized diagnosis and treatment efficiency.In short,this work plays an active role in the computer-aided diagnosis and treatment of cerebrovascular diseases. |
