A voxel-level approach to brain age prediction: A method to assess regional brain aging

TL;DR

The paper tackles regional brain aging by predicting voxel-level brain age from non-registered T1-weighted MRIs using a multitask U-Net that also performs global age prediction and GM/WM/CSF segmentation. It demonstrates significant voxel-level MAE improvements over a baseline across internal and external healthy test sets and reveals distinct regional aging patterns in dementia and Alzheimer's disease via PAD maps. Regional analyses, bias-correction, and an interpretability comparison show that voxel-level predictions yield quantitative, region-specific aging signals that align with known neurodegenerative patterns while offering new potential biomarkers. The work emphasizes robust cross-dataset performance, practical interpretability, and resources for reproducibility, including public code access.

Abstract

Brain aging is a regional phenomenon, a facet that remains relatively under-explored within the realm of brain age prediction research using machine learning methods. Voxel-level predictions can provide localized brain age estimates that can provide granular insights into the regional aging processes. This is essential to understand the differences in aging trajectories in healthy versus diseased subjects. In this work, a deep learning-based multitask model is proposed for voxel-level brain age prediction from T1-weighted magnetic resonance images. The proposed model outperforms the models existing in the literature and yields valuable clinical insights when applied to both healthy and diseased populations. Regional analysis is performed on the voxel-level brain age predictions to understand aging trajectories of known anatomical regions in the brain and show that there exist disparities in regional aging trajectories of healthy subjects compared to ones with underlying neurological disorders such as Dementia and more specifically, Alzheimer's disease. Our code is available at https://github.com/nehagianchandani/Voxel-level-brain-age-prediction.

Figures (11)

• Figure 1: Overview of the contributions of this article. There are 4 major contributions: (i) Proposal of a voxel-level brain age prediction model with initial validation on healthy subjects and subjects with dementia and Alzheimer's disease (ii) Ablation experiments were done to justify the use of a multitask architecture of the DL model with three-tasks over its two-task, and one-task counterparts. (iii) Regional analysis of predicted brain age by clustering voxel-level brain age predictions into known anatomical regions of the brain and (iv) An interpretability analysis where the proposed voxel-level approach to understanding regional aging trajectories is compared to traditional interpretability methods like Grad-CAM selvaraju2017grad, SmoothGrad smilkov2017smoothgrad, and Occlusion Sensitivity zeiler2014visualizing.
• Figure 2: The proposed multitask model used for voxel-level brain age prediction. The model follows a U-Net backbone with a downsampling and upsampling block to obtain the output at the same size as the input. It takes cropped MR patches of size $128\times128\times128$ as input to the model and produces three outputs, a voxel-level brain age prediction i.e. the main output as well as two secondary outputs, brain tissue segmentation of GM, WM and CSF, and global brain age prediction.
• Figure 3: (left) Cam-CAN training set follows a rough uniform data distribution, exposing the proposed model to samples of all ages. (right) This leads to bias-free predictions mostly, except for the extremities (ages 20-30 and ages 80-90). It can be observed that the predictions are closely aligned around the regression line for ages 30-80, with slight bias observed on the edges. A correction methodology can help correct the observed bias.
• Figure 4: Row 1 - PAD maps (raw, with no bias correction) based on the voxel-level difference between chronological and predicted age, Row 2 - adjusted PAD maps by subtracting the overall MAE of the brain volume from each voxel PAD value. Each row is plotted with an independent colormap based on the range of values observed in the samples plotted in that row. Minimum and maximum points on the color bar denote the minimum and maximum voxel PAD observed across the samples in the specific row. The data distribution plot beside the color bar in row 1 shows the distribution of PAD values across the entire healthy test set. Extended analysis of healthy PAD maps is described in gianchandani2023.
• Figure 5: Violin plots comparing the distribution of test results (voxel-level MAE) of different models. Each violin represents the probability density of data, with quartiles displayed as black lines and the median indicated by the white dot. It can be observed that for both the Cam-CAN and CC359 test sets, the median MAE is smaller for the proposed model compared to the baseline. All results are reported before bias correction.