(36) AI software demonstrates strong performance for screening of Tetralogy of Fallot and Truncus Arteriosus Communis

Congenital malformations are under-diagnosed during fetal ultrasound (US) exams. Artificial Intelligence (AI) could make the examination safer by drawing the practitioners eyes to suspicious images. Objective is to evaluate the feasibility to build an AI recognizing the overriding great vessel images of the Tetralogy of Fallot (ToF) or Truncus arteriosus communis (TAC).

Study Design:

An AI software (under development, not FDA approved) was trained on tens of thousands annotated US images from 30 major USA and European sites including the images of 102 patients with diagnosed ToF. No cases of TAC were used by the algorithm during the training phase. 

A first validation database is composed of 99 overriding aorta from ToF cases and 2153 normal images coming from these same centers. A second validation database was composed of 21 ToF and 7 TAC cases built from literature images between 2001 and 2021. Each exam was reviewed by one fetal US expert responsible for extracting all the images showing an overriding great vessel. The AI software was evaluated on its ability to accurately identify this sign.

Results:

On the 99 overriding aorta images of the first validation database, the AI software reached 94.95% sensitivity in raising an alert. On the literature images, respectively on ToF and TAC cases, the algorithm identified an overriding great vessel on resp. 36 and 12 images, reaching a sensitivity of resp. 90% and 92.3%. On the control group, 98.8% specificity was reached (Fig.1).

Conclusion:

This study demonstrated the feasibility in building a reliable alert system for screening of ToF along with the performance of such an algorithm in detecting pathologies non-seen during training (TAC) thanks to including well-known ultrasound semiology to guide the AI learning phase. Further tests on independent DBs from new centers are needed to better assess the AI software’s robustness.