Optimal Method for Fetal Brain Age Prediction Using Multiplanar Slices From Structural Magnetic Resonance Imaging

Abstract

The accurate prediction of fetal brain age using magnetic resonance imaging (MRI) may contribute to the identification of brain abnormalities and the risk of adverse developmental outcomes. This study aimed to propose a method for predicting fetal brain age using MRIs from 220 healthy fetuses between 15.9 and 38.7 weeks of gestational age (GA). We built a 2D single-channel convolutional neural network (CNN) with multiplanar MRI slices in different orthogonal planes without correction for interslice motion. In each fetus, multiple age predictions from different slices were generated, and the brain age was obtained using the mode that determined the most frequent value among the multiple predictions from the 2D single-channel CNN. We obtained a mean absolute error (MAE) of 0.125 weeks (0.875 days) between the GA and brain age across the fetuses. The use of multiplanar slices achieved significantly lower prediction error and its variance than the use of a single slice and a single MRI stack. Our 2D single-channel CNN with multiplanar slices yielded a significantly lower stack-wise MAE (0.304 weeks) than the 2D multi-channel (MAE = 0.979, p < 0.001) and 3D (MAE = 1.114, p < 0.001) CNNs. The saliency maps from our method indicated that the anatomical information describing the cortex and ventricles was the primary contributor to brain age prediction. With the application of the proposed method to external MRIs from 21 healthy fetuses, we obtained an MAE of 0.508 weeks. Based on the external MRIs, we found that the stack-wise MAE of the 2D single-channel CNN (0.743 weeks) was significantly lower than those of the 2D multi-channel (1.466 weeks, p < 0.001) and 3D (1.241 weeks, p < 0.001) CNNs. These results demonstrate that our method with multiplanar slices accurately predicts fetal brain age without the need for increased dimensionality or complex MRI preprocessing steps.

Fetal MRI by Robust Deep Generative Prior Reconstruction and Diffeomorphic Registration

“…We also observe consistently better results for all models when using the brain rather than the trunk ROI, with best figures 0.683 (0.950) provided again by DenseNet-201. These results compare favourably with those reported by [39], 0.751 (0.947), and [14], 0.767 (0.920), and unfavourably with those in [40], 0.508 (0.992). Further to this, with the exception of R 2 for poorly performing MobileNet-v2 and GoogLeNet, additional improvements are consistently observed for all models when combining volumetric brain and trunk information, a unique feature of the proposed technique, with MAE (R 2 ) 0.618 (0.958) for DenseNet-201.…”

“…We have shown that deep feature extraction using pre-trained models combined with correlation constrained linear regression provides accurate results for this task. Our results look competitive to existing methods [14], [39], [40], particularly when complementing brain features with trunk information, but there are differences in the cohorts considered. Most notably, existing methods use larger cohorts, including single-sequence data from 220 [40] and 289 subjects [39] and multi-sequence data from 764 subjects [14].…”

“…Our results look competitive to existing methods [14], [39], [40], particularly when complementing brain features with trunk information, but there are differences in the cohorts considered. Most notably, existing methods use larger cohorts, including single-sequence data from 220 [40] and 289 subjects [39] and multi-sequence data from 764 subjects [14]. Small sample sizes in our case precluded isolation of a subset of subjects for testing, a potentially important limitation when compared to [14] and [40].…”

“…Most notably, existing methods use larger cohorts, including single-sequence data from 220 [40] and 289 subjects [39] and multi-sequence data from 764 subjects [14]. Small sample sizes in our case precluded isolation of a subset of subjects for testing, a potentially important limitation when compared to [14] and [40].…”

“…To ensure comparisons are performed using the same FOV, we assess the performance of RGDSVR and DSVR for GA estimation using a trunk ROI corresponding to the FOV used as the target reconstruction volume in the experiments in [11]. As existing methods for fetal GA estimation from MRI [14], [39], [40] are based on brain data, we also test the GA estimation performance using a brain ROI from our reconstructions. In all cases we use the GA estimation method described in § II-H with 3D space spanned by slices from N r = 200 random volume reorientations.…”