Machine learning-based detection of fetal anatomical orientation in second trimester ultrasound images

To develop a machine learning-based method for detecting fetal anatomical orientation automatically using second-trimester fetal stomach and heart ultrasound image segmentation to facilitate automatic detection of situs abnormalities in future. Ultrasound images of the fetal stomach (227) and heart (142) were retrieved from a prospective study of routine ultrasound scans recorded on patients with normal pregnancies between 13-26 weeks of gestation at the John Radcliffe Hospital in England. The method combines segmentation and orientation of the fetal anatomy. Three machine learning models were trained to segment the fetal anatomy; the best model was selected by comparing the Dice coefficient, which compares the pixel-wise agreement between predicted and ground truth segmentation. The Target Registration Error (TRE) assessed the accuracy of fetal anatomy centroids, and probabilistic sensitivity assessed the accuracy of fetal anatomy orientation. The best model achieved a Dice score of 0.953 and 0.949 for segmenting the fetal stomach and heart, respectively. The median TRE for the stomach was 0.25 mm (min 0.01 to max 2.83), and for the heart was 0.39 mm (0.05 to 13.77). Assessment for orientation of true positives 137 (96.47%), false positives 3 (2.11%), and false negatives 2 (1.40%) showed a sensitivity of 98.5%. Our study demonstrates the feasibility of using a machine learning-based method to automatically detect fetal anatomical orientation in routine second-trimester ultrasound scans. This could be of interest in supporting the detection of heart malposition, which, in turn, increases the likelihood of congenital heart disease; or situs anomalies like Dextrocardia. Further research on automation for detecting anatomy orientation could simplify detecting anatomical orientation.