Abstract:

Objective Aortic stenosis (AS), a prevalent valvular disease, demands accurate diagnosis. Current methods, notably Doppler echocardiography, face limitations like dynamic image challenges and reliance on cardiologist experience. To assess aortic stenosis, measuring the LVOT diameter is critical, as a 1 mm difference can result in a 10% variation in stroke volume. Accurate Doppler beam alignment and LVOT VTI measurement are also essential to avoid errors. Our study, utilizing the TMED 2 dataset, introduces a novel artificial intelligence program for precise aortic stenosis diagnosis. By leveraging AI, we aim to overcome existing constraints and significantly enhance diagnostic accuracy.
Methods a novel method that involves using convolutional neural networks (CNNs), were used to grade AS based on various views of transthoracic echocardiography (TTE) images from the TMED 2 dataset. This innovative method aimed to take advantage of CNN’s abilities to recognize detailed patterns in echocardiographic data, making AS diagnosis more accurate. We evaluated the performance of our CNN models using confusion metrics and the area under the receiver operator curve (AUROC).
Results Our CNN networks were trained on a dataset comprising view_and_diagnosis_labeled_set, which included 599 studies from 577 unique patients (some with multiple studies on distinct days). For classification, we chose three classes: no aortic stenosis, aortic stenosis, and mild aortic stenosis. The detection of aortic stenosis achieved an accuracy of 85.74%. External validation using three views (PLAX, PSAX, and A4C) of outpatient transthoracic echocardiograms demonstrated effective screening for AS, yielding respective AUROCs of 0.81, 0.88, and 0.48.
Conclusion Our novel CNN-based approach achieved an 85,74% accuracy in AS detection using diverse views from the TMED 2 dataset. External validation on outpatient echocardiograms demonstrated robust screening capabilities, with AUROCs of 0.81, 0.88, and 0.48 for PLAX, PSAX, and A4C views, respectively. These promising results suggest the potential of AI in improving AS diagnosis for clinical applications. Moving forward, our future endeavors will focus on addressing data imbalances and detecting the view of images, in addition to assessing the severity of aortic stenosis, to further refine and optimize our diagnostic approach.

Key words: Aortic stenosis; Convolutional neural networks; Doppler echocardiography; Parasternal long-axis; Artificial intelligence.