Objective To evaluate the performance of a deep learning 1.5T cardiac Cine MR images in automatically quantification of the left ventricular (LV) function of patients with different cardiomyopathy.

Methods The cardiac MRI data of 393 subjects from March 2014 to November 2019 were retrospectively analyzed. The LV function of patients with hypertrophic cardiomyopathy (HCM group, n=125), patients with dilated cardiomyopathy (DCM group, n=133), and normal individuals (control group, n=135) was evaluated by manual and automatic measurements. The manual analysis was completed by two experienced physicians. After automatic measurement, the performance of convolutional neural network (CNN) was evaluated from aspects of left ventricular segmentation accuracy and LV functional parameter accuracy. Pearson correlation analysis, Bland-Altman analysis, and receiver operator curve (ROC) were used to evaluate the correlation and consistency between automatic and manual diagnosis for HCM and DCM.

Results When deep learning CNN was used to evaluate the LV function, the consistency with manual measurement was best in the HCM group, followed by the normal group, and the worst in the DCM group. The results of automatic and manual analysis of the four parameters of LV function in the HCM group had higher correlations (P < 0.01), and the correlations in the DCM group were weaker than those in the HCM group, especially ejection fraction and stroke volume. ROC curve analysis showed that the ejection fraction calculated by automatic segmentation has a certain diagnostic efficiency for DCM and HCM, with a sensitivity of 92.31% and 78.05%, and a specificity of 82.96% and 54.07%, respectively.

Conclusions Among different myocardial lesions, deep learning CNN-based cardiac function analysis may have different performances, which is worse in DCM and better in HCM, but has a greater diagnosis value for DCM than HCM.