Research On Left Ventricular Segmentation Based On Multi-scale Dilated Convolution And Adversarial Learning

The increasing mortality of cardiovascular diseases in recent years has caused widespread concern about cardiovascular diseases.Left ventricular segmentation based on cardiac magnetic resonance imaging can provide important cardiac function parameters,which play a key role in the diagnosis and treatment of cardiovascular diseases.Manual left ventricular segmentation requires not only professional personnel,but also the efficiency.Therefore,computer-based ancillary medical technology has emerged.The traditional automated left ventricular segmentation algorithm has shortcomings such as poor robustness,insufficient segmentation precision,too complicated algorithm operation and the need for professional participation.In these aspects,the left ventricular segmentation algorithm based on deep learning has great advantages.Due to the difference of patients and the angle of the photograph,there is a huge difference in the area of the closure area formed by the left ventricular membrane.Some current left ventricular segmentation algorithms based on deep learning are often difficult to effectively segment some special slices(such as the left ventricular end-systolic MRIs of patients with cardiac hypertrophy);on the other hand,the lack of labeled samples for training will affect the overall performance of the algorithms.In view of the above problems,this paper proposes a multi-scale dilated convolution and adversarial learning based left ventricular segmentation algorithm.The left ventricular segmentation algorithm based on multiscale dilated convolution uses the basic framework of a full convolutional neural network(FCN).Multiple cascaded multi-scale dilated convolution modules are used in the feature extraction(ie,coding)phase.Skip connection and dense upsampling convolutions are used in the segmentation mapping(ie,decoding)phase to connect to the final segmentation result.At the same time,in order to prevent the parameters of the model from being too large,this paper reasonably sets the number of convolution filters according to the requirements of the segmentation task and the characteristics of the magnetic resonance image.Finally,the results show that the results of the left ventricular segmentation algorithm based on multi-scale dilated convolution have improved compared with some excellent left ventricular segmentation algorithms in recent years.In order to further improve the performance of left ventricular segmentation with limited label samples,this paper proposes a left ventricular segmentation algorithm based on adversarial learning.It combines the idea of Generative Adversarial Network and "self-learning".The segmentation network and discriminator network are trained by adversarial learning.They achieve certain effects and then unlabeled samples are added for semi-supervised training.The final experimental results show that the left ventricular segmentation algorithm based on adversarial learning can further improve the accuracy and robustness when training with the same number of expert label samples.