491 ScanNav® audit: an AI-powered screening assistant for fetal anatomical ultrasound

Abstract

To develop and evaluate a real-time Artificial Intelligence (AI) based system to automatically keep track of acquired images; and check that the images conform to imaging protocol standards, in essence, replacing a human peer reviewer. We developed an AI system (ScanNav) which automatically (1) checks the completeness of the imaging record during fetal anomaly screening, ensuring all 19 required fetal views are recorded; and (2) assesses the quality of these images (in this case according to the guidelines of ISUOG). First, AI algorithms were trained on images manually evaluated by a pool of experienced sonologists and using state-of-the-art deep learning technology. The resulting algorithm was then assessed on a separate testing set; it was deemed correct if 2 or more (from a panel of 5) independent sonologists agreed with its decision. Due to the lack of expert agreement for “marginal” images, it was deemed appropriate to include as agreement such a 2:3 panel split. The system was developed on 479,322 anonymised images from 48,161 routine mid-trimester scans. Agreement between ScanNav and the sonologist panel was performed on an independent set of 38,840 images (4,284 scans). For scan completeness the mean (standard deviation) of agreement between ScanNav and the sonologists was 93.5% (±5.6%); for image quality it was 92.2% (±4.7%). The system processes 11 frames per second on a PC with an RTX4000 GPU. A real-time AI system was developed and extensively evaluated to automatically assess completeness and quality-check fetal ultrasound images against guidelines. Acting as a peer-reviewer it performs as well as an experienced sonologist. Within clinical practice such a system could ensure imaging records of scans are complete and conforming to a well-defined protocol. By improving the confidence of trainee and less experienced sonologists, the system has the potential to help the expansion of the sonologist workforce, while it can support expert sonologists by improving workflow.