Echocardiography is the most widely utilized cardiac imaging modality. It can diagnose a variety of heart conditions. Furthermore, echocardiography can provide quantitative traits of key parameters of cardiac structure and function, including ventricular remodeling, atrial remodeling, valvular function and arterial function. Echocardiography is acquired in a typical fashion, following specific views of the heart. Previous echocardiography AI approaches have been limited by lack of detailed input information, lack of a stepwise logical approach to cardiac diagnosis and quantification, and limited sample sizes. Through a collaboration between UGent and UPenn, we propose to overcome these barriers and develop an AI platform for automatic echocardiographic feature detection, quantification and diagnosis. We will utilize existing datasets that have been quantified and/or segmented, that have been labeled carefully regarding specific views and/or that have been interpreted by expert clinicians. These datasets include:

• ~750 clinical echocardiograms containing ~80,000 images or loops that have been carefully annotated regarding specific views through a detailed view classification nomenclature system designed specifically for this project.
• Several thousand echocardiograms from the Asklepios study in which flow envelopes have been traced in Matlab and ECG tracings have been extracted digitally already from existing images
• Clinical echocardiograms available at Penn Medicine (> 100,000 available) in which detailed specific diagnoses have been made by expert readers in a standardized structured fashion (rather than “free narrative” reporting).