Fetal Brain Age Estimation Using Feature-based Machine Learning and Image-based Deep Learning

Abstract

Fetal magnetic resonance imaging (MRI) has the potential to advance our understanding of human brain development by providing quantitative information of development in vivo. Among the quantitative information, accurate prediction of fetal brain age is important for evaluating brain health. This study describes methods to predict fetal brain age using a feature- based machine learning and image-based deep learning models. I propose a fully convolutional neural network for the automatic segmentation of cortical plate (CP). Accurate segmentation of CP is crucial for providing the cortical volume and folding measures. The cortical volume and folding measure related to brain development thus I used the measures as a feature of regression. The proposed method obtained Dice coefficients of 0.907 ± 0.027 and 0.906 ± 0.031 as well as a mean surface distance error of 0.182 ± 0.058 mm and 0.185 ± 0.069 mm for the left and right, respectively. The accuracy of regression, which is sigmoid fitting using surface area, was mean absolute error of 0.721and 0.852 and coefficient of determination (R2) of 0.897 and 0.856 for the left and right. I also demonstrated that the proposed hybrid loss function and the combination of multi-view aggregation and test-time augmentation (TTA) significantly improved the CP segmentation accuracy. On the other hand, I built a 2D single-channel convolutional neural network with MRI slices. I directly obtained multiple age predictions from multiplanar slices in each fetus using 2D networks, and the mode of the multiple predictions yielded a mean absolute error of 0.125 weeks across fetuses. My 2D multiplanar single-channel network obtained significantly lower mean absolute error than 2D multi-channel and 3D volume networks. It also showed higher prediction accuracy compared with regression models using volume- and surface-based measures. Saliency maps from my method indicate that anatomical information of the cortex and ventricles were primary contributors to brain age prediction. Lastly, using the difference values of brain age prediction compare with chronological age, we showed the diagnostic performance of disorder groups. In summary, I developed two methods for fetal brain age prediction. One is the morphological feature-based regression approach. However, the features were calculated from automatic segmentation results using the proposed segmentation network. The other is the image-based deep learning model for fetal brain age. I evaluated the prediction accuracy of each two methods and analyzed the pros and cons of each method.