Cardic CT Image Registration And Segmentation Based On Convolutional Neural Network

The heart pumps blood in the circulatory system and it is one of the most important organs in the human body.But cardiovascular diseases seriously threaten human health.Computed Tomography(CT),one of the most common medical imaging methods,plays an indispensable role in the diagnosis and adjuvant treatment of various diseases.In clinical medicine,segmentation of cardiac CT images provides important help for the diagnosis and treatment of cardiovascular diseases.Artificial segmentation of cardiac CT images is a complex and time-consuming task.In the field of medical image processing,automatic and accurate segmentation of cardiac CT images is an important research direction.Segmentation technology based on multi-atlas registration is a common method of image segmentation.Spatial transformation from atlas to target image can be obtained by registration technology.Applying this transformation to the label of atals can result the predicted segmentation of the target image.Fusion of multiple predicted results can usually improve the accuracy of the final segmentation.However,multi-atlas registration technology is not real-time due to the high time complexity.At the same time,the fusion of multiple predicted results is also difficult.The rapid development of Convolutional Neural Network(CNN)and the emergence of Fully Convolutional Network(FCN)make it possible to segment medical images quickly and accurately.But the accuracy of segmentation algorithm based on convolutional neural network is easily affected by the quality of datasets.In order to make use of the advantages of multi-atlas registration and convolution neural network in cardiac CT image segmentation,a three-dimensional cardiac CT image segmentation algorithm based on convolution neural network and multi-atlas registration is proposed in this paper.In order to reduce the time complexity of traditional LDDMM registration model,we use full convolution network with UNet structure to generate the initial momentum,which is needed for registration.In addition,in order to improve the accuracy of multi-label fusion process,we use neural network instead of traditional similarity measure,uncertain number candidate label and semi-local label fusion algorithm.In order to evaluate the performance of the proposed cardiac CT image segmentation algorithm,we carried out a detailed and systematic experiment with analysis on the MM-WHS dataset for each step of the algorithm.At the same time,we use Dice,MeanIU,NCC,MI and other evaluation strategies.The experimental results show that the proposed cardiac CT image segmentation method,which is based on the combination of neural network and multi-template registration strategy,has high accuracy and practicability.