OC07.04: Artificial intelligence software demonstrates strong performance for screening of transposition of great arteries
Year: 2024
Type: conference-abstract
Abstract: Congenital malformations are under-diagnosed during fetal ultrasound (US) exams. Artificial Intelligence (AI) could provide a safety belt drawing the practitioner's eyes to suspicious images. The aim is to evaluate the feasibility to build an AI recognising the different ultrasound signs of the transposition of great arteries (TGA). An AI software under development (not approved) was trained on ten of thousands annotated US non-pathological images from major USA and European sites. This training database was enriched with 190 cases (21015 images) of diagnosed TGA. Non-pathological validation database included 400 images of standard fetal heart views of 100 patients from major US and European sites. Validation database of TGA cases was built from images of the literature (22 cases coming from 13 different countries between 2006 and 2018). The variety of geography, manufacturers and image quality ensure the robustness of the algorithm. Each exam was reviewed by one fetal US expert extracting the images showing one of these signs: only two vessels visible on 3 vessels view; ventriculo-arterial discordance and parallel track of the great arteries. The AI software was evaluated on its ability to raise an alert when at least one of these signs was visible on the image. In the training database of TGA cases, 1316 images with at least one sign visible on it were extracted. On these, the AI software reached 98% sensitivity in raising an alert. From the validation database of TGA cases, 65 images were extracted with at least one sign visible on which the software reached 88% sensitivity. The specificity was 97% on the control group. This study demonstrated the feasibility in building a reliable alert system for screening of TGA. Further tests on independent databases from new centres are needed to better assess the AI software's robustness. True clinical benefit of such an alert system still remains to be evaluated by further investigations.
DOI: 10.1002/uog.27765
Authors: R. Besson, G. Corda, N. Matevski, E. Martinez, E. Spaggiari, N. Fries, Y. Oyelese, A. Fishman, A.C. Combs, R. Chaoui, J. Stirnemann, Y. Ville
Language: English
Subfield: Health Informatics
Field: Medicine
Domain: Health Sciences
Open Access status: bronze
OpenAlex ID: W4402359633