334 - Brain age estimation from neonatal MRI using deep neural networks

Poster Number: 334
Publication Number: 334.442

Farzad Beizaee, École de technologie supérieure, Montreal, PQ, Canada; Vicente Enguix, Research center CHU Sainte-Justine, Montreal, PQ, Canada; Christian Desrosiers, École de technologie supérieure, Montreal, PQ, Canada; Jose Dolz, ETS Montreal, Montreal, PQ, Canada; Gregory A. Lodygensky, CHU Sainte-Justine Dept of Pediatrics, Montréal, PQ, Canada

Background: Several studies have documented the incredible brain growth during the last trimester of gestation in fetuses and in preterm infants. Deep neural networks (DNN) offer the possibility of both image analyses and large complex calculations required to build a fully automatic brain age estimation pipeline.

Objective: First, we aim at successfully segmenting neonatal brain MRIs into regional structures by using a state-of-the-art DNN. And secondly, we intend to predict the gestational age of preterm and term neonates from the segmented MRI images by using a regression model based on neural networks.

Design/Methods: We trained a DNN (3DUNet [Çiçek, Ö. et al. MICCAI 2016]) on 340 healthy MRI scans between 29 and 44 weeks from the DHCP [Makropoulo,s A. et al. TMI 2014] using Draw-EM contours from 85 regional structures as ground truth. We then compared its performance to Draw-EM in terms of overlapping by using the well-known Dice Similarity Coefficient (DSC). Last, we trained a regression neural network to predict the neonates’ age from the brain regional structures’ size derived from both Draw-EM and UNet contours. The proposed method overview is shown in figure 1. To train the age prediction model, 80% of the data was used as the training set and the remaining 20% was considered as the validation set. Two metrics were used to assess the performance of the model: coefficient of determination (R2), which measures how well the regression predictions approximate the real data points, and Mean Absolute Error (MAE), which provides the average prediction error. Experiments were repeated 5 times for different training and validation folds and final scores were the averages of the 5 trials.

Results:
Individual MRIs were segmented with 3DUNet in a matter of a few seconds, whose mean DSC was 0.91, indicating a high overlap with the Draw-EM contours.
The regression model based on Draw-EM contours predicted neonates’ age with a MAE of 0.63±0.04 weeks and a R2 of 0.91±0.03. Surprisingly, the regression model based on 3DUNet segmentations predicted brain age more accurately, with a MAE of 0.60±0.04 weeks and R2 of 0.93±0.01. Furthermore, the method performs well for both males and females with MAE of 0.59±0.05 and 0.73±0.09 weeks, respectively.
Conclusion(s): Deep neural networks can be used to successfully segment the brain into 85 regional structures. Furthermore, the results indicate that neonates’ scan age can be predicted accurately from their brain structures’ size with an average error of just 4 days. We plan to add additional neonates’ brain MRIs as it will further improve the accuracy of the method.
Farzad Beizaee's CVFarzad Beizaee.pdf