Right Ventricular Image Segmentation Based On Convolutional Neural Network

Cardiovascular disease is one of several important diseases that threaten human health.Many families are sick and their families are broken and their wives are scattered.Therefore,early detection and timely treatment are very important for patients.Usually,the clinician will derive clinical parameters such as ventricular volume,ejection fraction,and myocardial mass based on the segmented image of the ventricle.With these parameters,the patient’s cardiac function can be quantitatively analyzed.Because the right ventricle structure is more complex than the left ventricle,and the right ventricular wall is thinner than the left ventricular wall,making the right ventricle easily mixed with the surrounding tissue,right ventricular segmentation has been difficult,and the literature for right ventricular segmentation has been compared.Less,so this article mainly studies the right ventricular segmentation problem.The traditional right ventricular image segmentation method requires the guidance of experienced doctors.The segmentation process is time-consuming and laborious,and the segmentation effect is also related to the doctor’s experience,which makes it difficult for the segmentation method to have good generalization ability under the premise of high precision.Therefore,it is very important to propose a right ventricular segmentation algorithm with high segmentation precision,fast speed and strong generalization ability.After analyzing the shortcomings of the existing right ventricular segmentation algorithms at home and abroad,this paper proposes two improved networks based on Unet networks and densely connected networks.Among them,the improved network based on Unet solves the problem that the network experience field is small and the network input and output image size is inconsistent by using the extended convolution of samepadding.In addition,according to the size of the experimental image,an upsampling layer and a downsampling layer are removed,which reduces the amount of computation of the network.In order to improve the training speed of the network,a batch normalization layer is added to the network.Considering that the number of pixels in the right ventricle is small in the proportion of the whole image,the loss function of the pixel cross entropy cannot evaluate the segmentation effect as a whole,so the loss function of the soft Dyce coefficient is used instead of the loss function of the pixel cross entropy.Although the improved Unet network segmentation accuracy has been improved but it has not achieved the expected results,a 12-layer full convolutional neural network is constructed with the idea of densely connected networks,and the extended convolutional layer used in the previous experiments is The batch normalization layer is also applied to the network.The experimental results show that the improved dense connection network training parameters are 0.2M less than the 1.9M of the Unet network and 3.6M of the Unet improved network,and the split Dice score is 0.91 higher than the 0.83 of the Unet network and 0.86 of the improved Unet network.The robustness of the algorithm is much better than the traditional algorithm.