Deep Learning-based Approach For Automatic Left Ventricle Function Assessment

In recent years,with the continuous deepening of deep learning algorithm research,deep learning has become an important tool in the field of medical auxiliary diagnosis.The field of medical imaging has attracted more and more attention from researchers.Echocardiography is a kind of low-cost,non-invasive cardiac medical imaging,which is often used in clinical for cardiac function evaluation,cardiac diagnosis,etc.Cardiac function evaluation is a very important step in clinical diagnosis and the basis for subsequent disease diagnosis.At present,many researches have realized the auxiliary diagnosis of echocardiography through deep learning algorithms,such as cardiac view classification,ventricular atrial segmentation,disease classification and cardiac function evaluation,but most of them are limited by training cost,inference speed and inference accuracy,resulting in many methods cannot be applied to practice.In this paper,a method of automatic evaluation of cardiac function is proposed,which constructs a complete end-to-end cardiac parameter measurement system by combining a variety of deep learning algorithms,and provides the possibility of auxiliary diagnosis for ultrasound physicians through the platform deployment of web applications.The cardiac function evaluation process proposed in this paper mainly includes cardiac perspective classification,automatic recognition of ED（End diastolic）and ES（End systolic）,automatic noise reduction of ED and ES frames,and left ventricular segmentation and result quantification,this paper aims to design and optimize the overall process of cardiac function evaluation to achieve a fast and accurate inference process,the main research contents include:In the model training stage,this paper proposal the idea of image classification,proposes a network structure based on Siamese neural network to achieve automatic recognition of ED and ES,and greatly compresses the number of network layers and parameters of the model by optimizing the model feature extraction module,and improves the model inference accuracy and speed.In addition,in this paper,an echocardiogram noise reduction model based on Auto Encoder is proposed,which improves the subsequent left ventricular segmentation accuracy of ED and ES recognition,and then obtains more accurate left ventricular quantification indicators.In the model deployment stage,this paper improves the inference accuracy of ED and ES recognition models through knowledge distillation method,and on this basis,the model pruning method reduces the amount of model parameters and further improves the inference speed.The final experimental results show that the identification accuracy of the ED and ES identification methods proposed in this paper is higher than that of the latest methods.In addition,the results of left ventricular segmentation were improved after image noise reduction,and the correlation coefficient between the final cardiac function evaluation result and the corresponding clinical data was 0.73,which is a high correlation,and indicates the effectiveness of this method.The cardiac function evaluation platform in this paper provides end-to-end assessment function,and supports acceleration of the CPU,and the total inference time is only about 1s in the experimental environment of this paper.