Estimating brain age based on a uniform healthy population with deep learning and structural MRI

Abstract

Numerous studies have established that estimated brain age, as derived from statistical models trained on healthy populations, constitutes a valuable biomarker that is predictive of cognitive decline and various neurological diseases. In this work, we curate a large-scale heterogeneous dataset (N = 10,158, age range 18 - 97) of structural brain MRIs in a healthy population from multiple publicly-available sources, upon which we train a deep learning model for brain age estimation. The availability of the large-scale dataset enables a more uniform age distribution across adult life-span for effective age estimation with no bias toward certain age groups. We demonstrate that the age estimation accuracy, evaluated with mean absolute error (MAE) and correlation coefficient (r), outperforms previously reported methods in both a hold-out test set reflective of the custom population (MAE = 4.06 years, r = 0.970) and an independent life-span evaluation dataset (MAE = 4.21 years, r = 0.960) on which a previous study has evaluated. We further demonstrate the utility of the estimated age in life-span aging analysis of cognitive functions. Furthermore, we conduct extensive ablation tests and employ feature-attribution techniques to analyze which regions contribute the most predictive value, demonstrating the prominence of the frontal lobe as well as pattern shift across life-span.

In summary, we achieve age estimation performance comparable to previous studies but with a more heterogenous dataset confirming the efficacy of deep learning and the added utility of training with data both in larger number and more uniformly distributed than in previous studies. We demonstrate the regional contribution to our brain age predictions through multiple routes and confirm the association of divergence between estimated and chronological brain age with neuropsychological measures.

1. Introduction

Age estimation is the task of estimating an individual’s age based on a set of other covariates. In particular, a large body of research focuses on the task of predicting brain age based on imaging studies[1, 2, 3]. In addition to its utility in studying the aging process itself, estimated age, as derived from models trained on a healthy population, has emerged as a useful biomarker for diseases[4]. Specifically, the divergence of one’s estimated age from chronological age has been associated with various diseases, especially those thought to mimic an advanced age state[5, 6].

Typically, models for estimating brain age are trained on datasets representing a normal aging population free of obvious disease. When subsequently applied on (possibly abnormal) participants, predicted brain age has been linked to education and self-reported physical activity [7] and has been utilized to characterize diseases including Alzheimer’s Disease (AD) [8], schizophrenia [9], traumatic brain injury [10], etc., where deviation from normal aging trajectory occurs alongside disease state. Related to this, healthy individuals are generally recruited such that their age overlap with the age range of the disease being studied; that is, healthy controls for schizophrenia studies will be younger than those for AD. When creating a selection of structural neuroimaging it is likely that the source data will oversample certain age ranges.

There are three components necessary to determine brain age with this paradigm: 1) covariates: the measured brain characteristics which serve as inputs to the model, 2) dataset: the precise cohort of normal aging participants upon which the model is trained, 3) model: the precise algorithms used to estimate the brain age given the available covariates.

While brain characteristics can be derived via many methods, neuroimaging is the most common and comprehensive way to characterize the “brain state” in vivo. Within neuroimaging, past studies have addressed EEG (Electroencephalogram) [11], DTI (diffusion tensor imaging) [12], and resting state BOLD fMRI (blood-oxygen-level dependent functional MRI) [13], which reflect different physiological brain measures. However, T1-weighted (T1w) structural MRI, which reveals features of the underlying anatomical characteristics of the brain, including gray and white matter delineation and gyral and sulcal patterns, is the most common modality in brain age estimation research. Imaging-derived neuroanatomical characteristics are biomarkers particularly sensitive to the aging process [14]. Practically, as one most widely available and standardized neuroimaging modalities, T1w structural MRI can be easily acquired for a large population. And within structural MRI domain, studies have used derived summary variables such as regional volumes or thickness estimation [15], requiring additional software processing to generate such values, and also unprocessed MRI scans [16].

We believe that, for these tools to be broadly useful, they must be trained on adequately diverse datasets that reflect the diversity of the populations on which the model might potentially be deployed. This range of population can include location, scanner, protocol, recruitment criteria and age. In this study, we propose using a composite dataset aggregated from several publicly available multi-center neuroimaging datasets, representing a diverse healthy population. This healthy study population is both larger in scale than those investigated by most previous age estimation studies and specifically designed to enable training age estimation models, and their evaluation, with an approximately uniform age distribution across the adult life span.

Lastly, given a study population and selected brain characteristics, the age is estimated based on statistical machine learning techniques. Another way to formulate the age estimation problem is to extract generalizable features from the brain that best capture the chronological age of a person provided the individual is undergoing a typical aging process that is present in a general healthy population. Numerous traditional machine learning methods have been proposed for age estimation including relevance vector machines [17, 18], Gaussian processes [19, 20], random forests [21], hidden Markov models [22], and non-negative matrix factorization [23].

More modern deep learning based methods are especially well suited to this task provided enough training data, and have been previously applied by [16] who demonstrated favorable performance. Interpretability is a critical aspect of any deep neural network based method. We seek not only to achieve predictive performance, but also to derive insights by understanding which features are most predictive. In this work, we explore regional contributions in the regression task by considering both feature ablation experiments and an activation map based post hoc interpretation method.

In summary, we utilize a 3D deep convolutional neural network based regression model to estimate age using T1w structural MRI volumes from a diverse multi-study population that is sampled with even age distribution across adult age span. We seek to extend and improve upon existing age-prediction studies by evaluating uniform and non-uniform age distributions in training and by performing analyses examining the regional contribution of prediction. This model performs well in both the hold-out test set from the same population and also in an independent life-span evaluation dataset. The deviance of the estimated age from chronological age is linked to neuropsychological and neuromorphometric measures. We further demonstrate the involvement of the frontal lobe in brain age estimation, and the pattern variability across life-span.

2. Method

In this section, we first describe the population and the experimental setup used in this study. We then describe the MRI pre-processing steps and the convolutional neural network used to estimate age. Next, given our learned model for estimating brain age, we present analysis that associates the divergence of the estimated age from the chronological brain age with neuropsychological and neuromorphometric measures. Finally, we describe multiple modes of regionality analysis for identifying which regions are most predictive. These consist of both ablation experiments and an extension of the class activation mapping method for interpreting feature importance in deep networks to the regression setting.

2.1. Study Population

It is necessary to build an adequate neuroimaging dataset for age prediction in the full adult life span, especially given study recruitment criteria the dataset is not necessarily evenly distributed across age. Increased participation in open data consortia and imaging datasets greatly facilitates the collection of such data. In this work, we collect more than 30,000 T1w MRI scans from multiple open neuroimaging datasets. All scans were acquired for studies approved by local institutional review board, research ethics committee, or human investigations committees. The list of the datasets used in this study with the full names and sources are listed in supplementary Table S2. Among those, we only include participants with clear indication of normal neurological evaluations contingent on the individual data providers criteria for a subject that is considered free of disease. Specifically, we exclude participants with any neurological or psychiatric disease, and also participants with no available diagnosis label. We also chose 18 as the minimum age to cover adulthood and also to avoid the neurodevelopmental stage. This results in 10,158 MRI sessions from 6,142 unique participants, the statistics are summarized in Table 1.

Table 1:

Dataset information

Dataset	Nsessions	Nsubjects	age range	agesession mean ± std	gendersession M/F	gendersubject M/F
ADNI	2423	438	56.3 - 95.8	76.64 ± 6.05	1223/1192	217/221
AIBL	781	457	60 - 92	72.86 ± 6.53	359/421	199/258
NIFD	428	136	36.9 - 85.2	65.76 ± 7.66	185/243	59/77
IXI	561	561	20.0 - 86.3	48.67 ± 16.47	248/313	248/313
BGSP	1566	1566	19 - 35	21.54 ± 2.89	661/905	661/905
Cam-CAN	652	652	18 - 88	54.30 ± 18.59	322/330	322/330
OASIS-1	316	316	18 - 94	45.09 ± 23.90	119/197	119/197
OASIS-2	145	56	60 - 97	76.54 ± 7.99	52/93	19/37
SALD	467	467	19 - 80	45.07 ± 17.39	168/297	168/297
SLIM	972	561	18 - 28.5	20.68 ± 1.40	425/547	244/317
PPMI	130	74	30.6 - 81	60.95 ± 10.85	91/39	48/26
SchizConnect	742	567	18 - 70	35.36 ± 12.76	429/313	338/229
DLBS	301	301	20.6 - 89.1	53.62 ± 19.92	115/186	115/186
CoRR	1326	642	18 - 83	28.78 ± 12.39	689/637	343/299
total	10158	6142	18 - 97	47.92 ± 24.89	4764/5783	2780/3362

However, as shown in Figure 1A and B, where we illustrate the age distribution of the 10,158 sessions and 6,142 participants respectively, the age distribution of the population is poorly balanced. Although there are studies covering the full age span including normal aging studies Cam-CAN [24], IXI, SALD [25], DLBS [26], OASIS-1 [27], and consortium-based studies such as CoRR [28], SchizConnect [29], many of the public imaging studies either focused on age-related disease in the older adult population including ADNI, AIBL [30], OASIS-2 [31], PPMI [32], NIFD; or on young participants including BGSP [33], SLIM [34]. To alleviate the potential bias toward a certain age segment, we sought to balance the age distribution that we ultimately use in the training population.

Figure 1:

The age distribution of the study population. A) The age distribution of the raw dataset consisted of 10,158 MRI sessions; B) the age distribution of the dataset consisted of 6,142 unique participants; C) the age distribution of the evenly sampled dataset.

In this study, when constructing the dataset to be uniform across age-span, we adopt both ‘oversampling’ and ‘undersampling’. Briefly, we oversample participants from age ranges with fewer participants by including the longitudinal follow-up sessions from the same participants, which could be regarded as a natural augmentation. For age ranges with more participants, we only include one scan per subject to increase the variability of the sample. If that number is still above the minimum across age bins, we further undersample stratified on confounding factors including acquisition site and gender.

We stratify the populations into age bins and use the bin with the minimum number of participants as the basis number. The age bins used in this study are [18, 20), [20, 25), [25, 30), … , [85, 90), [90, 100).

One interesting observation is that the [35, 40), [40, 45) are the two age bins with fewest number of participants, so we use the number of participants in this age segment as the base level and allow repeated scans from same participants. The other age bins having multiple scans per subject are the two age bins at the tail end: [85, 90) and [90, 100], because of the relative lower number in these two age bins. We undersample the participants in all other bins.

The final dataset consists of 2,852 MRI sessions from 2,694 participants covering age range 18-97, with the mean age 54.34 years old, standard deviation 21.16 years old. The age distribution of the evenly-sampled adult age span dataset is shown in Figure 1C.

2.2. Experimental setup

We design training, validation and test sets of participants for model training and evaluation. We build a validation set reflecting of general MRI scan distribution and population distribution. Similarly, a test set representative of the same population as the training and validation sets is important for evaluation.

Given this uniform age-distributed dataset we described in the section above, we perform stratified split based on acquisition sites and gender within each age bin: 8/10 as training set, 1/10 as validation set, 1/10 as test set, ensuring non-overlapping participants and similar distribution of age, gender and acquisition sites in the training, validation and test sets.

We also evaluate our model by testing it on an independent test set: Cam-CAN. Aiming to study the normal aging process through the adult life-span, Cam-CAN provided even age distribution across adult life-span and was previously used as an independent testing sample in another age estimation study [19]. We note that the trained model should be applicable to any similarly acquired structural neuroimaging within this age range. However, the result on an independent sample, which usually reflects a homogeneous population or acquisition setting, might be over- or under-optimistic.

Additionally, we also perform a test-retest experiment using an independent dataset of three participants scanned 40 times in 30 days [35] to evaluate the reproducibility of the model.

2.3. MRI Preprocessing

We apply basic pre-processing steps including nonparametric nonuniform intensity normalization (N3) based bias field correction [36], brain extraction using FreeSurfer [37], and affine registration to the 1mm³ isotropic MNI152 brain template with trilinear interpolation using FSL FLIRT [38]. The dimension of the 3D volume is 182 × 218 × 182 (LR×AP×SI).

All the preprocessed scans were checked by a well-trained reviewer with visual inspection. Scans having obvious and severe MRI artifacts, brain extraction failure or poor registration were excluded.

2.4. Convolutional neural network

We use a three-dimensional convolutional neural network (3D-CNN) regression model for age estimation, with similar architecture as the 3D-CNN classification model for Alzheimer’s disease classification used in [39]. We follow a general CNN architecture similar to the VGG classification architecture [40] with multiple interleaved convolutional blocks and max pooling layers and an increasing number of features along the depth. We replaced all 2D operations with 3D operations and included one fully-connected layer to reduce the number of parameters. We use convolutional layers with kernel size of 3 × 3 × 3, rectified linear unit (ReLU) as the activation function, and batch normalization (BN) layers before the activation functions. We flatten the output from the last convolutional layer and feed into a fully-connected (FC) layer. And the final activation is a linear operation instead of a softmax in classification tasks. We include weight l2 regularization to prevent overfitting with a factor of 1.0. The algorithm was optimized using Adam algorithm with mean absolute error (MAE) loss function, and with a batch size of 5. The initial learning rate was tuned in the range from 1e – 4 to 1e – 6 including [1e-4, 5e-5, 2e-5, 1e-5, 5e-6, 2e-6, 1e-6] and was set at 2e-5.

We implemented the algorithm using Keras and TensorFlow. An illustration of the framework is shown in Figure 2. In this study, we use five (N in Figure 2) stages. The feature dimension of the first layer is 16 and increases by a factor of 2 in each subsequent stage. The optimal model is selected as the model with the lowest validation MAE.

Figure 2:

The convolution neural network architecture. The inputs are 3D brain volumes. Each cube represents one 3D feature map. The size of the cube reflects the spatial dimension of the feature map, the number of the cubes reflects the number of feature maps (channel dimension) at a specific depth. The blue arrow denotes 3D convolution operation with rectified linear unit (ReLU), the green arrow denotes 3D convolution followed by batch normalization (BN) and ReLU, the yellow arrow denotes the max pooling operation. The basic unit enclosed in the bracket is repeated N = 5 times with increasing number of features and decreasing spatial dimension. The final convolutional output is flattened and fed into one fully-connected (FC) layer with linear output (red arrow), generating the final brain age prediction.

2.5. Comparison with model trained on unbalanced dataset

We also trained the model using all scans from unique participants (N = 6,142, Figure 1B). We applied the trained model on the independent Cam-CAN dataset and studied the distribution of MAE over chronological age groups.

Besides, a simple way to potentially correct the imbalance without adjusting the sampling of the dataset is re-weighting the samples, specifically, we assign different weights to different sample, with the weights in proportional to the inverse of the frequency of specific age segments. We note that while importance weighting is a principled statistical technique with well-known effects in traditional statistical models, its impact on the learned predictors in deep learning algorithms are a subject of active inquiry [41].

2.6. Neuropsychological and morphometric associations

To test the utility of the estimated age in studying cognitive functions across adult life-span, we evaluate the association between the summary scores of Benton face recognition test (BFRT) and the estimated age in Cam-CAN dataset.

Specifically, we use the signed difference of the estimated age and chronological age to reflect the deviance of individual brain age from their chronological age, and we refer to this value as age_diff hereinafter. The BFRT is a commonly used neuropsychological instrument that can be easily and reliably administered in adult patients to test baseline visual memory and perception [42]. We adopted the SubScore-1, SubScore-2, TotalScore (SubScore-1 + SubScore-2) as dependent variables in individual linear regression models incorporating gender, chronological age, age_diff, and the interaction of chronological age and age_diff:

$$\text{Score}\sim {\beta }_{\text{age}}age+{\beta }_{{\text{age}}_{\text{diff}}}{\text{age}}_{\text{diff}}+{\beta }_{{\text{age}}_{\text{diff}}\times \text{age}}\text{age}\times {\text{age}}_{\text{diff}}+{\beta }_{\text{gender}}\text{gender}$$

Additionally, we evaluate the association between age_diff with cortical thickness generated using FreeSurfer [43] by performing a partial correlation with gender and chronological age as covariates. FreeSurfer parcellates the cortex into 68 cortical regions.

2.7. Age activation map

Class activation mapping [44, 45] is a commonly-used method for interpreting the classification using CNNs and has been previously used in CNN based medical image analysis [46] to marry potential disease pathology with classification findings. In this work, we use the idea of a class activation map in a regression setting by highlighting the small-valued gradient in grad-CAM framework. We use functions in the keras-vis package (https://github.com/raghakot/keras-vis/). We generate the average activation map within each age group to investigate the age-specific pattern of underlying substrates for age estimation.

2.8. Slice based age estimation

Besides the post hoc saliency map based activation map method, we also propose ablation analyses methods focusing on part of the input data. We apply serial 2D CNNs for age estimation with the input being three consecutive slices along each axis. By doing so, we take into consideration the inter-subject alignment precision (i.e. not using just one slice) and also the similarity among different channels (i.e. not using five slices). The network architecture of the 2D CNN is the identical to the 3D CNN architecture described in the previous section with the 3D operations replaced with the corresponding 2D operations. We report the predictive performance on various sets of 2D slices to analyze predictive importance.

2.9. Lobe based age estimation

Besides sliced based age estimation, we propose using another more anatomically-informed way to study the regionality through ablation experiments at the lobar level. The individual lobe masks were generated following a previous study [39]. The ages are estimated using each lobe individually.

3. Results

3.1. Age prediction

In the hold-out test set, whose instances are distributed identically as the training and validation set, our model achieves an MAE of 4.06 years and correlation coefficient r = 0.970.

An independent normal aging life-span dataset—Cam-CAN—was tested in a previous brain age study [19]. In that study, when the Cam-CAN data were pre-processed with the optimal parameters selected from the independent training sample, their proposed model achieved an MAE of 6.08 years and correlation coefficient r = 0.929. We tested out our model in Cam-CAN study processing the T1w MRI images using the same pipeline as the other samples. The relationship between the estimated age and chronological age in Cam-CAN is shown in Figure 3. The correlation coefficient r is 0.960, MAE is 4.21 years, which outperforms the result in the previous study [19]. We also observe two obvious “outliers” among the 652 participants tested, further investigation is needed to pinpoint potential sources of error, either methodological, constitutional or true poor estimation. The results demonstrate our proposed model achieves accurate estimation in all age segments.

Figure 3:

The estimated age versus chronological age in an independent test set of adult life-span.

3.2. Reproducibility

We evaluate the reproducibility of the algorithm in test-retest acquisitions. We show the results in Table 2 and Figure 4, observing that there is a difference in the estimated age and actual reported chronological age that is consistent over the sessions with approximately a 1 year standard deviation.

Table 2:

Reproducibility experiment result

subject	actual age	predicted age mean	predicted age std
subj-1	26	25.19	1.07
subj-2	31	33.02	1.14
subj-3	30	27.06	0.81

Figure 4:

The distribution of predicted ages in test-retest scans.

3.3. Comparison with results using nonuniform dataset

We compare the results using nonuniform dataset with the MAE performance in Cam-CAN dataset. Using the nonuniform dataset achieves comparable overall MAE (4.27 years) as the balanced data. Re-weighting the samples helps slightly improves the MAE (4.17 years) than the balanced dataset despite using many more scans. Additionally, we observe the MAE using the nonuniform dataset is not evenly distributed across life-span: MAE is lower in the young age with more abundant data, as shown in Figure 5 (B). This could introduce potential bias in life-span studies. Using sample re-weighting (Figure 5 (C)) alleviates the problem and using balanced dataset generates generally even distribution across age-span (Figure 5 (A)).

Figure 5:

Distribution of MAE across life-span. (Left) Age estimated using the balanced dataset. Each step in the red line indicate the MAE in that age group, the black dashed line indicates the overall MAE. (Middle) Age estimated using the nonuniform dataset. (Right) Age estimated using the nonuniform dataset but with sample re-weighting.

3.4. Neuropsychological and neuromorphometric association

The association of the BFRT scores with the difference in the estimated age and the chronological age and its interaction with chronological age are summarized in Table 3, with the p-value of the regression corresponding to age, age_diff and age_diff × age.

Table 3:

Association with Benton face recognition scores

Score	age		age diff		age diff × age
	β	p	β	p	β	p
SubScore-1	−2.51e-3	3.01e-6	−0.0127	0.0297	2.62e-4	0.0102
SubScore-2	−0.0605	1.24e-32	−0.157	2.90e-3	2.31e-3	0.0119
TotalScore	−0.0630	5.07e-34	−0.170	1.54e-3	2.57e-3	5.88e-3

The association of the cortical thickness measures with the age_diff is illustrated in Figure 6. The thickness of cortical regions are significantly associated with the age_diff. In addition, out of the 68 regions measured, 51 regions have a stronger correlation with the estimated age than the chronological age. This is expected as the age estimated through structural MRI image is in principle more coupled to structural phenotypes.

Figure 6:

The partial correlation coefficients of age_diff and cortical thickness measures. The red dashed line indicates α = 0.05 under multiple comparison of N=34 regions.

3.5. Age activation map

We expect the anatomical patterns characterizing different age groups to be different throughout lifetime but are consistent within a local age range. Thus we generate and illustrate the age activation maps every 5 years, in the same way as preparing the dataset. The 3D iso-surfaces of the average age activation maps are shown in Fig. 7. The average age activation maps overlaid on the MNI152 template are shown in Fig. 8 (Left). To accommodate the anatomical differences in different age groups, we also generated average T1w image within each age group, and overlaid the corresponding age activation map, as shown in Fig. 8 (Right).

Figure 7:

The 3D iso-surfaces (0.8) of the age activation maps at different age groups.

Figure 8:

The age activation maps at different age groups overlaid on the (Left) MNI152 template, and (Right) average T1w image within each age group, both with threshold at 0.8. Each slice is identified in the plane to identify the voxel of highest intensity.

3.6. Slice based age estimation

The age estimation performance using 2D MRI slabs sliced at different coordinate planes is shown in Figure 9. The slices with the best performance are also illustrated.

Figure 9:

MRI 2D slice based age estimation. (Top row) The mean absolute error (MAE) of the estimated age on the test set using 2D slices at different locations, the red lines indicate the location with lowest MAE. MAEs larger than 10 are capped to 10 for illustration purpose. (Bottom row) The illustration of slices at the red lines in the top row from the MNI152 template and the corresponding segmentation (the colors follow the FreeSurfer color lookup table).

3.7. Lobe based age estimation

The MAE trained using different lobes are shown in Table 4, where we show the best estimation performance achieved through frontal lobe. And temporal lobe achieved marginally lower performance.

Table 4:

Lobe based age estimation

Lobe	Frontal	Temporal	Parietal	Occipital	Cerebellum	Whole-brain
MAE	5.33	5.81	6.37	7.66	6.20	4.06

4. Discussion

In this study, a large heterogeneous dataset of structural neuroimaging across the adult lifespan was aggregated from multiple publicly available data sources. A deep convolutional neural network was developed, trained and applied to predict the chronological age from the subject’s scan.

This uniformly distributed dataset is able to achieve estimation without appreciable bias towards a certain age group, while maintaining training efficiency compared with training on an nonuniform dataset. While this study aims to examine the aging process across the adult life-span, other studies aiming to examine the role of aging in other conditions would likely benefit from other specific training dataset inclusion criteria. For example, for studies of autism and prodromal psychosis where age of onset skews younger, studies would require the inclusion of participants below 18, and participants in middle and old age would likely be less informative.

The regionality analysis in this study reveals patterns of neuroanatomical contributions of normal aging. All analyses provide evidence for the prominence of frontal regions in all epochs of age estimation in the adult lifespan. Frontal regions have been implicated in normal aging through both neuropsychological studies and neuroimaging studies [47, 48, 49]. In addition, the pattern shifts reflected in the age activation map based analysis imply a frontal lobe-focused locus of age-related structural changes. We found a negative correlation between several frontal lobe stucture’s cortical thickness measures and difference in age. This may be due to gray and white matter having differing contributions to the age-prediction model.

Neuropsychological evaluations targeting different cognitive domains and brain regions might help reveal the underlying complexity in the aging-process. Although these saliency maps are not well understood, it may be possible they reflect patterns of intensity changes driven by either shrinkage or expansion of ventricles, or differences in the boundary of gray or white matter. More studies are needed to explain and understand these contributions.

Our analysis revealed the association between the divergence of estimated age from chronological age and BFRT performance. BFRT scores are strongly correlated to actual age, in that performance decreases with increased age. While deviations in estimated age versus BFRT reflect two processes that may independently deviate through different mechanisms from a subject’s age, it is useful to have a secondary independent measurement available upon which to compare chronological age. Deviations in BFRT and the difference in estimated age might reflect data points to study to improve the accuracy of the model, or provide insight into why young appearing brains perform worse, or older appearing brains perform better than expected. This suggests the potential utility of the estimated age at normal aging evaluation, in complement to other cognitive test and neuroimaging based measures. The utility relates to an open question of the aging process. It still requires further validation how a scan’s predicted deviance would reveal an individual’s brain health status or even trigger clinical evaluations, since while inter-subject differences in normal aging process exist, they are not well understood. Moreover, it is unclear whether this inherent variability increases with age or stays constant throughout the life-span.

One consideration is that the set of participants with no current pathological symptoms might have disease at asymptomatic prodromal state that have yet to manifest in the clinical evaluation, such as pre-clinical AD. Such work would require follow-up data to preclude disease for a given period of time such that they remain stable as controls. Otherwise, one would be unable to determine the difference between prodromal Alzheimer’s disease and MCI. However, this was not available for all participants, and in a group of otherwise healthy individuals, it is likely that some healthy individuals do harbor occult disease. Additionally, participants were considered normal using criteria germane to the disease of focus in different studies.

Another consideration in uniform sampling is the tails, specifically the youngest and oldest group. For this study to focus on adult age prediction, there are practical limitations including the reduced imaging data available for very old individuals and that this study doesn’t include individuals under 18 years of age. Future studies should examine on stratification procedures keeping in mind the difficulty of obtaining indivdual scans and adolescent development.

One consideration is the leave-N-out validation scheme, in lieu of a cross-validation. Although computationally much more taxing, future studies using it may help explain variance in age estimation differences. While ideally one would want to include both methods, for large imaging studies it is important to consider when developing and evaluating a large-scale neuroimaging model.

We do not report that this model achieve the most superior age-prediction results using a deep learning or similar framework using T1-weighted structural images compared to existing studies, but that it in fact examines the contribution of age distribution in a training set. This is understudied in age-prediction when using large data-sets because there is limited diversity to the images chosen. In addition, there is added diversity of the studies from which the model was trained, and the protocols and acquisitions may vary across both site and study. This in fact benefits this model compared to individual studies of unified aging because the variance of acquisition differences are included in the training, compared to training data where a more cohesive protocol is used, alongside uniform age separation. Such inclusion without appreciable loss of accuracy supports the use of this type of model on a wider variety of baseline scans.

5. Conclusion

This study has demonstrated the high accuracy of an age estimation framework using routine structural neuroimaging and a deep convolutional neural network, trained on a large-scale healthy population with uniform age distribution across adult age span. While there is considerable interest and value in a highly accurate estimation model that can be applied to a wide variety of structural MRI protocols, this study also demonstrates patterns of the regional contributions of aging across the lifespan using multiple methods, agnostic of tissue classes, structure delineation or surface parcellation. This study provides a framework that may have value in evaluating meaningful quantitative and visual biomarkers for both aging and broad neuroimaging applications. Further studies may examine the estimated age and the regional contributions to see if they may be useful in understanding one’s age divergence in the context of a particular disease or condition’s pathology.