Learning deep architectures for the Interpretation of Fetal Echocardiography

Grant: 408PED /2020, PN-III-P2-2.1-PED-2019-2227

Contracting authority: UEFISCDI

Grant value: 600.000 RON

Coordinating institution: Romanian Institute of Science and Technology

Partner: University of Medicine and Pharmacy of Craiova

Implementation period: 01.11.2020-31.10.2022

Abstract:

Medicine is one of the main and most important stages where deep learning has been playing a successful role during the last years. The project team has also targeted the creation of such methods for medical images, resulting in a framework that supports both classification and instance segmentation. A further challenging enhancement would be to include models for a more complex task, i.e. to analyze medical data from video scans. A good choice for such a problem is fetal ultrasound, where there is a moving patient-in-patient. An even more stimulating scenario is to additionally examine a moving organ, e.g. the heart. Fetal cardiac examination is at the same time one of the crucial assessments during pregnancy towards a prompt detection of possible congenital heart disease. In a fetal echocardiography, the physicians look for the absence, distortion or improper positioning of different heart components per se or in relation to one another, as well as for an incorrect blood flow in the heart. Apart from the complexity of the examination, there are also the issues of commonly performing scan only for the second trimester (when termination is no longer a choice), the large amount of echocardiograms from screening and the risk of a poor detection rate in the absence of a protocol or of an extensive experience. In this respect, the available deep learning framework enhanced with models to recognize structure and behavior of first and second trimester fetal heart from video analysis can implicitly reach the identification of distinguishing features and delineation of heart components and blood flow. The models can indicate to the clinician the views of interest with the presence/absence of certain characteristics of importance in the diagnosis. The superior framework will provide a proof of concept of a virtual assistant for the obstetrician in observing the heart abnormalities present in the videos and support an early, cost-effective diagnosis of a serious disease.

Obtained results

The LIFE project aimed at an early analysis of first and second trimester fetal echocardiography with the support of artificial intelligence (deep learning).

In the first trimester of pregnancy, the heart development is just at its beginning. Four key planes must be recognized as present —the aortic flow, atrioventricular flows, arches, and pulmonary and ductal flows — on a tiny heart with fetal movement and shadows present on sonographic examination.

In the second trimester of pregnancy, the heart is in full structural development. From the key plans, several elements must be identified - the septum, atria, ventricles, open and closed valves of the aortic-left ventricular outflow tract, pulmonary artery, superior vena cava etc.

Artificial intelligence, through deep learning, can learn from physician-annotated video frames what the planes and structures of interest are, and can then autonomously and efficiently recognize key features directly on a new scan. It thus proves to be a useful support tool in the early maternal-fetal examination. A demo for first trimester can be seen at https://youtu.be/QzFmaAKTqlc, while for the second at https://youtu.be/riT30C1A9Rw.