Deep Learning and Multiplex Networks for Accurate Modeling of Brain Age

Abstract

Recent works have extensively investigated the possibility to predict brain aging from T1-weighted MRI brain scans. The main purposes of these studies are the investigation of subject-specific aging mechanisms and the development of accurate models for age prediction. Deviations between predicted and chronological age are known to occur in several neurodegenerative diseases; as a consequence, reaching higher levels of age prediction accuracy is of paramount importance to develop diagnostic tools. In this work, we propose a novel complex network model for brain based on segmenting T1-weighted MRI scans in rectangular boxes, called patches, and measuring pairwise similarities using Pearson's correlation to define a subject-specific network. We fed a deep neural network with nodal metrics, evaluating both the intensity and the uniformity of connections, to predict subjects' ages. Our model reaches high accuracies which compare favorably with state-of-the-art approaches. We observe that the complex relationships involved in this brain description cannot be accurately modeled with standard machine learning approaches, such as Ridge and Lasso regression, Random Forest, and Support Vector Machines, instead a deep neural network has to be used.

Keywords: age prediction; aging; brain; deep learning; lifespan; machine learning; multiplex networks; structural MRI.

Figure 1

After dividing the brain into left and right hemispheres, each hemisphere is divided in 300 patches. This operation is performed for each subject in the cohort, after registration, thus each patch is expected to roughly contain the same anatomical district and analogous distributions of white matter, gray matter, and cerebrospinal fluid.

Figure 2

Age shapes brain networks by modifying the spatial distribution of white matter, gray matter, and cerebrospinal fluid. Accordingly, nodal metrics, such as strength and inverse participation, allow the detection and quantification of these age-related changes.

Figure 3

A schematic representation of our deep neural network. It consisted of four hidden layers composed of 200, 100, 50, and 20 neurons.

Figure 4

Random Forest algorithm consists of two main phases: (1) data sample is bootstrapped N times, N is the number of trees and the first parameter this algorithm needs to be set; (2) when growing a tree j, at each node splitting i a random set of features f_ij is sampled; the features selected change at every split, but the number of sampled features, the second parameter to be set, remains constant.

Figure 5

From left to right, histogram of cross-validation results: MAE, RMSE, and Pearson's correlation ρ.

Figure 6

Overall scatter plot of chronological age (x-axis) and predicted age (y-axis) and the specific four age ranges (right panel): 7 ≤ Age < 20 (A), 20 ≤ Age < 40 (B), 40 ≤ Age < 60 (C), 60 ≤ Age < 80 (D).

Figure 7

We evaluated the regression metrics MAE, RMSE, and correlation by randomly sampling a varying percentage of subjects from the whole cohort, from 10 to 100%, and reported the results of 100 ten-fold cross-validations.

Figure 8

This figure shows the patches related to the first most important features along 5 axial planes of the MNI 152 template. On the bottom right, a 3D representation of the patches on the reference space is reported, as well.