Study Of Cerebrovascular Reconstruction In Magnetic Resonance Angiography

Magnetic Resonance Angiography(MRA)is a routine examination.Its accurate 3D reconstruction of cerebrovascular is an important auxiliary method for the diagnosis and treatment of cerebrovascular diseases,which has received extensive attention in the research field.With the rapid development of computer technology,cerebrovascular segmentation,as an important prerequisite for 3D reconstruction,has been developed into deep learning as the main method,which is exploring in the direction of lightweight models and high-precision segmentaion.Based on the need for precise treatment and limited MRA datasets,it presents a larger challenge to cerebrovascular segmentation.Aiming at the limitations of the existing methods,this paper carried out researches on the 3D reconstruction method of cerebrovascular based on deep learning.It runs through the design of deep learning models,the improvement of training methods,and the fusion of probabilistic prediction.Its main contents are as follows(1)Aiming at the information redundancy of convolutional neural networks in feature space,this paper proposes a convolutional neural network for adaptive cerebrovascular texture enhancement.Compared with regular convolutional kernels,the proposed Gabor convolutional kernel has a consistent physical structure and can be ’plug-and-play’ in any convolutional neural network.Based on the multi-channel parallel computing of convolutional neural network,the Gabor kernel can satisfy the multi-angle and multi-scale enhancement of vessel texture features,and update Gabor parameters adaptively in the back propagation.Its adaptability reduces the complexity of the feature space,so that a lightweight network can be designed based on the Gabor kernel.Three vessel datasets were selected to test the Gabor kernel.The results show that the proposed convolutional neural network can effectively improve the efficiency of vessel segmentation,and the total parameters is less than 1%of the classic convolutional neural network.It is conducive to lightweight deployment.(2)To improve the feature expression ability of segmentation model in small sample datasets,an adversarial training method based on segmentation adversarial neural network is proposed.A segmentation adversarial neural network consists of a discriminator and segmentation model.In the discriminator,a dual-branch discriminator with feature reuse is proposed.It performs true and false probabilistic predictions for the manual segmentation and the model prediction.The discriminator is optimized towards distinguishing them,and uses a gradient loss to constrain the variation caused by the change of manual segmentation.In the segmentation model,the penalty passed from the discriminator is combined in the regular supervised penalty,and an attempt is made to ’trick’ the discriminator.Two models are trained together in adversarial manner to improve the performance of the segmentation model.The experiments show that the performance is significantly improved using adversarial training optimization.(3)In view of the complexity of manual segmentation,a semi-supervised learning is proposed for cerebrovascular segmentation in sparsely labeled dataset.The semi-supervised learning framework adopts a two-thread parallel method to implement supervised training based on a small amount of labeled dataset.It adds random disturbance to alter the distribution of unlabeled datasets,thereby establishing new mappings through a consistency penalty.To prevent the loss value from oscillating during the training process,the labeled dataset is traversed as a training epoch,and the consistency loss is randomly calculated in each step.Also,to strengthen the spatial correlation of cerebrovascular images,a new network is proposed based on Transformer,which improves the fusion way and strengthens the parsing efficiency of spatial information.In experiments,a total of 10 groups of deep learning models were compared to verify the effectiveness of the proposed model.The simulating sparsely labeled dataset tested the semi-supervised learning framework.The results show that the proposed framework can reduce the dependence on labeled datasets.(4)Using the probabilistic prediction from different models,a cerebrovascular information fusion method based on evidence theory is proposed.The hypothesis space quantifies the fuzzy information from predictions with low confidence and improve the prediction accuracy through confidence inference.To further improve the missing information of cerebrovascular due to weak signal and artifacts in the imaging process,we extracted the skeleton and performed the topological analysis.Based on the curvature consistency,the relationship between the end nodes of the fractured vessel is established,and the geometric shape simulation complements the vessel information.The experiments verify the effectiveness of evidencetheoretic information fusion,and assess the accuracy of completion information using multi-modal MRA.