EP01.34: Towards AI‐enabled navigation of the second trimester anomaly scan

Using sonographer eye tracking during ultrasound, we have previously developed a saliency prediction model (PURFECT) for second trimester anomaly scanning. Using a deep convolutional network architecture, the model generates a saliency heat map predicting the visual gaze of an experienced sonographer when given a real-time ultrasound image (see figure 1). In this study, we determine whether this AI-model could be used to support navigation in order to guide operators to an optimum plane of visualisation during the second trimester scan. Second trimester scans undertaken prospectively using the PURFECT system were recorded at the John Radcliffe Hospital (University of Oxford) between February and October 2023. We included low-risk, singleton pregnancies at 18-23 weeks. For each image of the fetal head circumference (HC) and trans-cerebellar view (TCV), the PURFECT saliency heat map was compared to manually segmented areas, and the area of overlap calculated. The PURFECT precision (overlap/segmented area) and recall (overlap/PURFECT heat-map) were calculated. Twenty recorded videos were randomly selected. For HC, the PURFECT heatmap focal point was within a segmented area in 75% of images, precision was 44.7%, recall was 14.4%. For TCV, the PURFECT heatmap focal point was within a segmented area in 95% of images, precision was 80.2%, recall was 17.8%. PURFECT was more likely to highlight anatomical structures when they were well magnified and situated in the centre of the screen. In some cases, the predicted visual gaze of sonographers highlights standard structures of anatomical interest, but this alone is insufficient to guide operators reliably towards optimal image ascertainment. Prior anatomical knowledge is likely needed to further develop AI-enabled navigation. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.