Arq Bras Cardiol: Imagem cardiovasc. 2024; 37(4): e20240095

Artificial Intelligence in Echocardiography: The Future of Precision Diagnosis

Maria Estefânia Bosco

Otto

DOI: 10.36660/abcimg.20240095i

AI and the Systolic Function

Ejection fraction (EF) is an essential parameter used to assess the systolic function; however, it has been historically subjected to interobserver variability and the level of experience of the examiner. AI-based algorithms present solutions for EF, providing automated and reproducible calculations.^3–5 A pioneering study published by He et al. evaluated a deep learning (DL) system comprised of convolutional neural networks (CNNs), called EchoNet-Dynamic flow, and compared EF assessments performed by AI with measurements performed by experienced sonographers. The reference cardiologists who checked these results were unable to differentiate between the sonographer and AI. Furthermore, the results demonstrated that the assessment by AI was not inferior to those performed by experts, with a difference of less than 5% in variability. Other advantages of this system were the speed of measurement and the possibility of averaging up to five heartbeats, thus improving the accuracy of the result in atrial fibrillation.

Another parameter widely used to assess systolic function in echocardiography is the Global Longitudinal Strain (GLS) analysis, which not only aids in the diagnosis of cardiac function, but also reveals characteristic patterns of certain diseases, such as amyloidosis. The GLS technique itself requires machine learning (ML) training of the equipment to identify acoustic window incidences and outline edges. However, AP techniques by CNR were able to assess GLS more quickly and with less variability when compared to a commercially established workstation used in echocardiography laboratories. Recently, Kwan et al. evaluated a CNR system used to assess GLS with images produced by equipment from two different companies, concluding that, in addition to having less variability and good reproducibility, this system was able to identify GLS patterns and differentiate between athlete’s hypertrophy, amyloidosis, and hypertrophic cardiomyopathy. However, it important to understand that the quality of the images obtained for both GLS and EF assessments must be good, a factor that cannot yet be achieved by AI.

[…]