Recognition of abnormal changes in echocardiographic videos by an artificial intelligence assisted diagnosis model based on 3D CNN

Abstract

Objective: To investigate the diagnostic efficiency and clinical application value of an artificial intelligence-assisted diagnosis model based on a three-dimensional convolutional neural network (3D CNN) on echocardiographic videos of patients with hypertensive heart disease, chronic renal failure (CRF) and hypothyroidism with cardiac involvement. Methods: This study is a retrospective study. The patients with hypertensive heart disease, CRF and hypothyroidism with cardiac involvement, who admitted in Henan Provincial People's Hospital from April 2019 to October 2021, were enrolled. Patients were divided into hypertension group, CRF group, and hypothyroidism group. Additionally, a simple random sampling method was used to select control healthy individuals, who underwent physical examination at the same period. The echocardiographic video data of enrolled participants were analyzed. The video data in each group was divided into a training set and an independent testing set in a ratio of 5 to 1. The temporal and spatial characteristics of videos were extracted using an inflated 3D convolutional network (I3D). The artificial intelligence assisted diagnosis model was trained and tested. There was no case overlapped between the training and validation sets. A model was established according to cases or videos based on video data from 3 different views (single apical four chamber (A4C) view, single parasternal left ventricular long-axis (PLAX) view and all views). The statistical analysis of diagnostic performance was completed to calculate sensitivity, specificity and area under the ROC curve (AUC). The time required for the artificial intelligence and ultrasound physicians to process cases was compared. Results: A total of 730 subjects aged (41.9±12.7) years were enrolled, including 362 males (49.6%), and 17 703 videos were collected. There were 212 cases in the hypertensive group, 210 cases in the CRF group, 105 cases in the hypothyroidism group, and 203 cases in the normal control group. The diagnostic performance of the model predicted by cases based on single PLAX view and all views data was excellent: (1) in the hypertensive group, the sensitivity, specificity and AUC of models based on all views data were 97%, 89% and 0.93, respectively, while those of models based on a single PLAX view were 94%, 95%, and 0.94, respectively; (2) in the CRF group, the sensitivity, specificity and AUC of models based on all views data were 97%, 95% and 0.96, respectively, while those of models based on a single PLAX view were 97%, 89%, and 0.93, respectively; (3) in the hypothyroidism group, the sensitivity, specificity and AUC of models based on all views data were 64%, 100% and 0.82, respectively, while those of models based on a single PLAX view were 82%, 89%, and 0.86, respectively. The time required for the 3D CNN model to measure and analyze the echocardiographic videos of each subject was significantly shorter than that for the ultrasound physicians ((23.96±6.65)s vs. (958.25±266.17)s, P<0.001). Conclusions: The artificial intelligence assisted diagnosis model based on 3D CNN can extract the dynamic temporal and spatial characteristics of echocardiographic videos jointly, and quickly and efficiently identify hypertensive heart disease and cardiac changes caused by CRF and hypothyroidism.