Study Of Full Myocardial Segmentation Algorithm Based On Deep Learning Methods

In recent years,cardiovascular diseases have become No.1 for worldwide death caused by noninfectious diseases.Myocardial segmentation plays an important role in cardiovascular disease diagnosis.However,in current clinical applications,accurate myocardium segmentation in CMR images depends on annotations from professional doctors,which costs a lot of time.Therefore,an automated myocardium segmentation with high accuracy is an appealing research goal.Most traditional image segmentation algorithms require manual participation and are difficult to achieve high segmentation accuracy.Challenges for the whole myocardium segmentation includes the indistinct boundary between myocardial tissue and the surrounding environment and the irregular shape of the right myocardium.Deep learning has recently achieved excellent performance in computer vision and image processing,and is becoming the mainstream of medical image processing.After analyzing the traditional image segmentation algorithm,we propose two automated segmentation of whole myocardium based on deep learning algorithms.The main contribution of the thesis includes:1,We propose an automated end-to-end myocardium segmentation model DDensenet.This model has the advantage of quickly expanding the receptive field without deepening networks,and therefore keeping more details.Experimental results in Kaggle cardiovascular image dataset demonstrated that our algorithm DDenseNet has acchived an DSC value of 87.27%while keeping its small volume(0.78M).2,By analyzing the advantages and disadvantages of U-Net,which is the state-of-art model in medical image segmentation,we found that the receptive field of U-Net has only 68 pixels in the deepest layer when an input image size is 256×256.This will result in the loss of the spatial information of the whole image in training.To solve this problem,we propose a RDU-Net model.It increases the receptive field without increasing the network parameters by reducing the downsampling module and adding the expansion convolution,and uses the residual mechanism to obtain more image features.Experimental results in Kaggle cardiovascular image dataset show that it achieves a DSC value of 90.61%,which is 8%higher than U-Net.Both models can be extended to medical image segmentation in other tissues,such as lung nodule segmentation,lung contour segmentation,etc..Our models have potential clinical application value.