Fetal Ultrasound Standard Plane Extraction using Orthogonal Triple-slice Deep Reinforcement Learning Agent

Using ultrasound for fetal anatomical survey and fetal growth monitoring can be challenging and tedious as sonographers need to manually search for a set of standard planes (SPs) using a 2D ultrasound probe. A desirable alternative is using a 3D ultrasound device, either hand-held or wearable, to capture large field-of-view volumetric images and apply an image analysis algorithm to automatically extract the target SPs. Prior work has been conducted to formulate this problem as iteratively moving a 6-degree-of-freedom 2D resampling plane toward the target viewing pose. However, views with insufficient anatomical information can lead to incorrect actions thus poor results for SP extraction.In this work, we propose to extend the 6-degree-of-freedom plane agent and leverage the two other resampling views orthogonal to the original plane to incorporate more context information for adaptive action prediction based on the most informative view. Experiments have been conducted on a preliminary clinical ultrasound dataset and the results show that with adaptive view selection, our algorithm can extract fetal biparietal diameter plane with an average plane localization error of 7.09 mm and 8.01 deg, comparing to the error of 25.35 mm and 35.15 deg when using a single view.