Manual versus automated segmentation for the computed assessment of the fetal face from ultrasound scans: preliminary results from a pilot study.

Objectives In order to develop computerised analysis of the fetal face, specific facial ‘landmarks’ must be identified and delineated on 3D ultrasound volumes. To date, this has been done by a time consuming manual process which would preclude this technology from widespread clinical use. We sought to assess the feasibility of an automated segmentation technique for quantitative assessment of fetal face three-dimensional ultrasound volumes. Methods 3D ultrasound fetal face volumes from 13 normal and 7 abnormal cases, collected for clinical indications at 29+4 (25-36+1) weeks and acquired in the midline sagittal view using a Voluson E8 system (GE Healthcare) equipped with low frequency probe, were used in this study. Manual segmentation was performed by tracing the fetal face surface in Mimics (Materialise, Belgium). Facial volumes were used to create an automatic segmentation tool as previously developed. A leave-one-out approach was used to test the automatedmethod and the overlapping between the two reconstructions was computed by means of Dice scores. Results The automatic segmentation tool successfully segmented the 20 meshes. Dice scores suggested satisfactory performance, with a mean overlapping of 75% between manual and automatic segmentations (range 0.61-0.85) (figure). The main facial features could be correctly represented in all cases. Conclusions This study demonstrates the feasibility of an automated approach for the reconstruction and quantification of facial shapes obtained from ultrasound 3D volumes. If our results were confirmed prospectively, automated segmentation might represent a fast and reliable tool for the reconstruction of ultrasound facial volumes acquired antenatally.