OC05.01: An AI system to support sonologists during fetal ultrasound anomaly screening

Abstract

We present a real-time AI system which supports a sonologist by automatically keeping track of acquired images and by checking that the images follow the Fetal Anomaly Screening Program (FASP) standards, in essence, replacing a human peer reviewer. We have developed an automatic AI system which 1) checks the completeness of images in an anomaly scan (scan completeness) and 2) assesses the quality of the acquired images against a clinical protocol (image quality). The AI algorithms are based on state-of-the-art deep learning technology and have been developed and evaluated on a large real-world dataset of routine anomaly scans (over 280,000 images). All ultrasound images were first manually evaluated by a pool of experienced sonologists. 75% of the images were used to train the system while 25% were used to assess the system performance. The automatic AI system performs as well as an experienced sonologist when assessing anomaly scan images: thus, when scan completeness was assessed by a peer reviewer and the AI system, agreement was 95.6% between them, and this was similar to the agreement between two peer reviewers (95.7%). In addition, the overall agreement on assessment of image quality was 84.2% between a peer and the AI system, while it was 84.0% between two peers. We have developed a real-time AI system to categorise and quality assure fetal anomaly images to support a sonologist during screening. The AI system performs as well as an experienced sonologist in assessing anomaly scan images. Use of such AI technology during scanning has the potential to reduce scanning times, support trainee sonologists, ensure completeness of scans, increase the adherence to clinical acquisition protocol, and thus further raise confidence in the use of ultrasound in this challenging clinical environment.

Ähnliche Arbeiten