OC24.02: Automated image analysis to estimate gestational age between 18 and 34 weeks using a single 3D brain volume dataset

Abstract

Ultrasound (US) estimates gestational age (GA) accurately up to ∼24 weeks, but women in low-income settings often present later in pregnancy and with uncertain dates. We hypothesized that GA can be estimated after 24 weeks by analysing the developing fetal brain using an automated algorithm requiring a single 3D US volume dataset. Standardized 3D volumes of the fetal brain were acquired in the axial plane as part of the INTERGROWTH-21st Project. A supervised machine-learning algorithm was developed with data from 18 - 33 + 6 weeks. Random regression forest analyses were applied on a training set of 150 volumes to provide a representation of the brain to ‘learn’ which features were best able to discriminate GA. Testing was performed on a prospectively acquired independent validation sample (n = 150). Model performance was assessed by comparing the modelled (GA: model) with the actual (GA: truth) GA, which was based on the last menstrual period and corroborated by measuring the first trimester crown–rump length. Systematic differences (SD) and 95% limits of agreement (LOA) were evaluated and heat-maps generated to illustrate GA discriminating regions of the brain. The model successfully analysed 138/150 validation volumes to predict GA: model. The SD for GA: model compared to GA:truth was +0.8 days; the 95% LOA were -10.92 to 10.52 days respectively. The thalamus and Sylvian fissure were regions with good GA discriminating features. Obtaining GA estimates using automated analysis of single 3D fetal brain volumes is feasible. Preliminary clinical testing suggests the model is stable. Estimates of model error at 34 weeks are consistent with conventional parameters at ∼22 weeks. Further testing using a range of clinical phenotypes is underway.