Cardiomyopathy recognition system based on convolution and long short-term memory neural network

Abstract

The invention relates to a cardiomyopathy recognition system based on a convolution and long short-term memory neural network, and belongs to the technical field of medical image analysis. The system comprises the following modules: 1, a heart nuclear magnetic picture preprocessing module; 2, a neural network architecture module; 3, a mobile terminal model MobileNet module; and 4, a classification loss function module containing prior information. According to the system, variants of a deep convolutional neural network and a convolutional long-short-term memory neural network are mainly combined, anatomical structures such as left ventricle myocardial internal and external membranes do not need to be manually segmented, classification training is performed by giving enough data sets based on certain priori knowledge, and finally, fitting determination of model parameters is realized; therefore, the heart nuclear magnetic data can be improved to complete the diagnosis accuracy of the fertile heart disease, the dilated heart disease and the normal cardiac muscle, and the processing and diagnosis time of the clinical heart magnetic resonance image can be shortened through the breakthrough of the technology.

Description

technical field [0001] The invention belongs to the technical field of medical image analysis, and relates to a cardiomyopathy recognition system based on convolution and long-short-term memory neural network. Background technique [0002] In the early stage of atrial magnetic resonance analysis and diagnosis, clinicians mainly relied on subjective experience and used manual segmentation software (such as: 3D slicer, ITK-SNAP, etc.) to divide each part of the heart and make corresponding diagnosis. With the popularity of machine learning, many machine learning solutions have also begun to be applied to the field of cardiac diagnosis. For example, in addition to the patient's height and weight, Khened et al. also used 9 features extracted from the segmentation map. Based on these features, they trained a random forest classifier with 100 trees for prediction. Wolterink et al. extracted 14 features (including 12 extracted from the segmentation map and the patient's height and...

AI Technical Summary

Problems solved by technology

However, it is undeniable that this algorithm is difficult to implement for large-scale training samples, and still needs to segment the image to extract information features, and the manual segmentation of the atrium takes a long time and is easily affected by subjective factors

Images