1049: Fully automated placental volume quantification from 3DUS for prediction of small-for-gestational age infants

Early placental volume (PV) is associated with small-for-gestational age infants born under the 10thor 5thcentiles (SGA10 and SGA5). However, 3DUS requires time-intensive, manual or semi-automated segmentation. For PV to become useful in a clinical setting, a fully automated segmentation method is needed. 3DUS images obtained at 11-14 weeks were put through our novel, fully automated pipeline (Figure 1), which starts by training a convolutional neural network (CNN) from a disjoint set of 47 manually annotated 3DUS images. This model is used to segment the study images. The results are fed into a joint label fusion (JLF) step as initialization. The results are combined via a second-tier random forest (RF) model to obtain final segmentations. Fig. 1b shows typical segmentation results. The study images were also segmented using the semi-automated VOCAL (GE Healthcare) method. The PV obtained with either method was used to train a logistic regression classifier for SGA10 and SGA5 outcomes. We either use the raw PV, or the normalized placental quotient (PQ=PV/gestational age at the time of US). A 5-fold stratified cross-validation setup is used, with 10 different seeds per fold. 341 subjects were included, with 35 (10.3%) and 19 (5.6%) delivering SGA10 and SGA5 infants, respectively. The mean gestational age at delivery was 39±1.7wks. There was a high correlation between the PV values obtained by each method (R2=0.78); although the mean PV was lower for VOCAL compared to CNN/JLF (66.2±18.1cc vs. 81.3± 21.2cc; p<1e-5), consistent with our prior observation that VOCAL underestimates PV. Figure 2 shows the ROC curves for predicting SGA10 and SGA5 outcomes with our CNN/JLF method and with VOCAL, averaged over the folds. The differences in AUC did not reach statistical significance. Our fully automated segmentation method successfully yields PV measurements at 11-14 weeks that are significantly associated with SGA, but with no manual user input required. This paves the way for exploring such a tool in a clinical setting as part of a multivariable prediction model for risk stratification and patient counseling. Moreover, it can allow a more rigorous investigation of placental shape and morphometry as potentially relevant markers of placental development.View Large Image Figure ViewerDownload Hi-res image Download (PPT)