Инновационная медицина Кубани (Nov 2024)

Artificial Intelligence in Fetal Echocardiography

  • E. L. Bokerija,
  • N. E. Yannaeva,
  • A. N. Sencha,
  • P. V. Metelkin,
  • O. V. Yurchenko

DOI
https://doi.org/10.35401/2541-9897-2024-9-4-14-20
Journal volume & issue
Vol. 0, no. 4
pp. 14 – 20

Abstract

Read online

Background: Congenital heart diseases (CHD) are one of the most common birth defects, occurring in 5-9 per 1000 newborns. CHD are the second leading cause of infant mortality and account for 47% of all causes of death from birth defects.The main method for assessing the anatomy and function of the heart is 2-dimensional ultrasonography. Artificial intelligence (AI) technologies are great at recognizing images, thus facilitating quick scanning and analysis of visual information in order to speed up and simplify the diagnostic ultrasonography.All AI software for obstetrics use static images. In our study conducted at the National Medical Research Center for Obstetrics, Gynecology and Perinatology named after Academician V.I. Kulakov (Moscow, Russian Federation) in 2022-2023, we used video files including 1-5 standard heart views for each fetus.Objective: To create a data set for development of an AI tool that improves the quality of fetal CHD diagnosis and to develop an algorithm for examining the fetal heart using AI. Resulting medical reports could be either “normal” (correct structure of the heart; no sign of CHD) or “abnormal” (incorrect structure of the heart; CHD cannot be excluded; extended fetal echocardiography is recommended as soon as possible).Materials and methods: The examination was conducted at 18-21 weeks’ gestation. Each examination contained video files of 5 standard views of the heart per patient. Each view is at least 25 frames. Verification was performed by confirming/changing the diagnosis by a physician and confirming the diagnosis after birth.Conclusions: As a result, the task of determining zones of the fetal chest and heart was solved with an approximate accuracy of 98%; the task of classifying the heart view on the frame was solved with an approximate accuracy of 82%, and the task of determining the disease on the heart views was solved with an approximate accuracy of 77%.

Keywords