Research And Implementation Of Right Ventricular MRI Segmentation System Based On Deep Learning

Nowadays,cardiovascular disease has become the "killer" of human health.In the process of treatment and examination of cardiovascular diseases,doctors get physiological information of patients by cardiac MRI.Doctors can deduce specific clinical indicators to evaluate and analyze the function and vitality of the heart,so segmentation and recognition of organs MRI is extremely important.In the main organs and tissues of the heart,there are individual differences in the right ventricle,and there are also differences in different stages,so the task of right ventricle segmentation is difficult.However,due to the difference of image equipment itself,the output image’s technical parameters are different and the manual segmentation is subjective,which will lead to the reduction of segmentation accuracy.In the task of right ventricular segmentation,the traditional methods such as region growing and threshold limiting have limited application scope and low accuracy.The deep learning method is mainly aimed at the heart,and the performance still has room for improvement.Therefore,in this paper,we use the deep learning to carry out in-depth research on the right ventricular image segmentation.On this basis,we develop a robust and accurate computer-aided image segmentation system,which can reduce the pressure of medical image data processing and enhance the clinical application of medical image information.At the same time,it is of great significance to promote the early intervention and late treatment of cardiovascular diseases.Based on the open MRI data set,this paper mainly focuses on the right ventricular segmentation task.First of all,in order to solve the problem of insufficient training samples,this paper use traditional transformation including translation,amplification,rotation and flipping to enhance the dimensions of data set.Aiming at the problem of right ventricle segmentation,this paper uses YOLO V3 to detect and locate the target area of right ventricle,and extracts the region of interest from the original image as the input of the next segmentation model.Taking Fully Convolutional Networks and UNet model as the baseline segmentation models,the network structure of Dense UNet,which uses dense connection block to extract image feature information,is compared with baseline models,and the final result is obtained by using dense connection condition random field algorithm to process the segmentation result.In the experiment,we evaluate the performance of the above segmentation models in Dice and Jaccard coefficients.The use of dense connection block of dense UNET network can greatly reduce the parameters of network training to speed up the training speed and improve the segmentation performance of model tasks.Dense UNET model has a certain improvement in stability and accuracy.Dice and Jaccard coefficients are 0.81 and 0.86.Based on the above mentioned deep learning model,this paper designs functions and architecture of the right ventricular segmentation system and realizes the corresponding modules by using the flask framework in the windows system environment.The system finally realizes functions of detecting and displaying the segmented right ventricular image on the web,and supports the function of correcting the segmented image and annotation of the model online.