Volume-based complete automation for ultrasound fetal biometry: A pilot approach to assess feasibility, reliability, and perspectives.

Abstract

Background: Detection algorithms targeting anatomic landmarks in three-dimensional (3D) ultrasound (US) volume (three-dimensional US) appear to be a relevant and easy-to-implement option to address junior and occasional operators' difficulties in probe positioning for two-dimensional (2D) fetal biometry.

Objectives: This study assesses the feasibility of complete automation for fetal biometry and the resulting agreement with standard 2D (US) measurements. The secondary objectives were to assess the impact of software-driven measurement on image quality scoring, reproducibility, and agreement with human-driven measurements issued from the same volumes.

Methods: Datasets were collected from a consecutive sample of women attending standard US follow-up (singleton, 16-30 weeks of gestation). Each dataset contained 2D measurements for reference (head and abdomen circumference and femoral length) and 3D US volume acquisitions of the fetal head, abdomen, and thigh. Both algorithm-based and operator-based detection of the targeted plans and calipers positioning were applied to the 3D volumes to produce software-driven and human-driven measurements. The resulting 3D measurements were assessed for completion rates, image quality, and reproducibility.

Results: On 175 datasets collected, completion rates in achieving software-driven 3D measurements ranged between 94% (abdomen) and 100% (head). A modest weakening in quality (of uncertain clinical significance) was notable for the head and abdomen measurements. Compared to the 2D measurements, the software-driven tended to slightly overestimate the estimated fetal weight (EFW; e.g., 95% confidence interval ranging from 445 to 635 g for a 525 g-sized fetus at 22 weeks of gestation). The random error tended to be inflated for fetuses >700 g. Intra- and inter-operator reproducibility were appropriate (intraclass correlation coefficient intervals ranged from 0.8 to 0.99).

Conclusion: Complete automation of US biometry appears feasible and presents appropriate reproducibility and image quality scoring, but third-trimester biometry needs improvement. Before clinical implementation, it is time to assess the impact of point-of-care use on large populations.