From a deep learning model back to the brain - inferring morphological markers and their relation to aging

Deep convolutional neural networks (CNN) enabled a major leap in image processing tasks including brain imaging analysis. In this work, we present a Deep Learning framework for the prediction of chronological age from structural MRI scans of healthy subjects. Previous findings associate an overestimation of brain age with neurodegenerative disease and higher mortality rates. However, the importance of brain age prediction and its discrepancy from the corresponding chronological age go beyond serving as biomarkers for neurological disorders. Specifically, utilizing CNN analysis to identify and locate brain regions and structures that contribute to the prediction can shed light on the complex multivariate process of brain aging. Previous work examined methods to attribute pixel/voxel-wise contribution to the prediction in a single image, resulting in ‘explanation maps’ (EM) that were found noisy and unreliable. To address this problem, we developed a novel inference framework for combining these maps across subjects, thus creating a population-based rather than subject-specific map. We apply this method to a CNN ensemble trained on predicting subjects’ chronological age from raw anatomical T1 brain images of 10,176 healthy subjects, obtained from various open-source datasets. Evaluating the model on an untouched test set (n = 588) resulted in MAE of 3.07 years and a correlation between the chronological and predicted age of r=0.98. Using the inference method, we revealed that cavities containing CSF, previously found as general atrophy markers, had the highest contribution for age prediction in our model. These were followed by subcortical GM, WM, and finally cortical GM. Comparing these maps derived from different models within the ensemble allowed to assess differences and similarities in the brain regions utilized by the model. To validate our method, we showed that it substantially increases the replicability of the EM as a function of sample size. Moreover, benchmarking our results against a baseline of voxel-based morphometry (VBM) studies revealed a significant overlap. Finally, we demonstrate that the maps highlight brain regions whose volumetric variability contributed the most to the model prediction.