In total, 290 T1-weighted brain MRIs from 274 rhesus monkeys were used in this study. They include 172 male and 102 female monkeys, 1–30 years of age at the time of brain MRI acquisition, 126 from the Boston University (BU) NHP database, and 148 from the PRIMatE Data Exchange (PRIME-DE) database (Milham et al., 2018).

The monkeys at BU were obtained from the Emory National Primate Research Center, the New England National Primate Research Center, and World Wide Primates for other funded projects over 3 decades. Prior to entering the study, all monkeys received medical examinations that included serum chemistry and health screenings to ensure that monkeys did not have a history of malnutrition, diabetes, chronic illness, or any neurological diseases. All monkeys were housed in the Laboratory Animal Science Center of Boston University Medical Campus, an AAALAC-accredited facility. Experiments were conducted in accordance with the Guide for the Care and Use of Laboratory Animals from the National Institute of Health's Office of Laboratory Animal Welfare and were approved by the Boston University Institutional Animal Care and Use Committee (IACUC).

MRIs on BU monkeys were obtained as previously described (Koo et al., 2012, 2013). In brief, the monkeys were anesthetized with an initial dose of ketamine (10 mg/kg, IM) followed by administration of propofol IV (25 mg/kg, IV) using a syringe pump to maintain anesthesia throughout the session or with repeated doses of ketamine (10 mg/kg, IV) and valium (1 mg/kg, IV). In the MRI scanner, monkeys were held in a stereotactic MRI-compatible head holder designed to fit within the 8-channel phase array head coil in a fixed position preventing any inadvertent movement from respiration. Respiration and oxygenation levels were monitored, and body temperature was maintained throughout the imaging session. MRI scans were performed on an Intera (for 16 scans) or Achieva 3T whole-body MRI scanner (Philips Healthcare, Best, The Netherlands). T1-weighted images were acquired using a T1-weighted 3D-turbo field echo (TFE) sequence: repetition time (TR)/echo time (TE) = 7 ms/3 ms, flip angle = 8°, NEX = 6, inter shot delay = 2,800 ms, TFE = factor 200, voxel size: = 0.6 mm isotropic, and sagittal plane acquisition.

The PRIME-DE database combines NHP neuroimaging data sets across 22 sites. Details have been described in a 2018 article (Milham et al., 2018). Of them, this study included data from 15 sites that contain in vivo structural MRI data for rhesus monkeys, details are described in Supplementary Table S2.

Among the 290 T1-weighted MRI scans, 260 monkeys had only one scan session, 12 had two MRI sessions, and 2 had three MRI sessions (at least 6 months apart between each scan). These longitudinal scans are treated as independent data for brain age estimation. Multiple scans from human subjects were also used in human brain age prediction studies with dozens or lower hundreds of subjects, but our study included a much lower percentage (6%) of human subjects with two scans.

We collected 8,859 T1-weighted human brain MRIs (0–97 years of age, 45.2% male) from 13 public data sets, including ABIDE-I (Di Martino et al., 2014) (567 scans, age 6–57), BeijingEN¹(180 scans, age 17–28), CamCAN (Taylor et al., 2017) (442 scans, age 18–68), DLBS (Park et al., 2012) (315 scans, age 20–89), DutchDonders² (171 scans, age 18–32), IXI³ (556 scans, age 20–86), Narratives (Nastase et al., 2021) (335 scans, age 18–53), NIH-PD (Evans et al., 2006) (1,210 scans, age 0–23), KI-RS_Enhanced (Nooner et al., 2012) (2,032 scans, age 6–85), OASIS-3 (LaMontagne et al., 2018) (1,838 scans, age 42–97), SALD (Wei et al., 2018) (488 scans, age 19–80), Wayne (Daugherty and Raz, 2017) (606 scans, age 18–91), and YaleHires (Finn et al., 2015) (119 scans, age 18–59). These T1-weighted brain MRIs were all acquired at 1 isotropic mm voxel size or slightly higher resolution.

Following the human brain age estimation (He et al., 2021a,b, 2022), we performed identical minimal pre-processing to both monkey and human brain MRIs.

For human brain T1-weighted brain MRI, the pre-processing included N4 bias correction (Tustison et al., 2010), field of view normalization (Ou et al., 2018), multi-atlas skull stripping (MASS) (Doshi et al., 2013; Ou et al., 2015), and a deformable registration of the skull-stripped image into the SRI24 atlas space (Rohlfing et al., 2010) by the deformable registration via attribute matching and mutual-saliency weighting (DRAMMS) algorithm (Ou et al., 2011, 2014). The deformable registration split the T1-weighted human brain MRI into two channels, both in the atlas space: a registered intensity image channel for the contrast information and a tissue density map known as RAVENS map (regional analysis of volumes examined in normalized space) containing morphometry information (Davatzikos et al., 2001). See Figure 1B for these two channels for exemplar subjects. Our previous study demonstrated that splitting a single T1-weighted MRI into these two channels led to a higher brain age estimation accuracy (He et al., 2021b, 2022). Therefore, this study used both channels for age estimation.

Monkey T1-weighted brain MRIs underwent the AFNI pipeline (animalwarper tool) to generate skull-stripped images (Saad et al., 2009; Jung et al., 2021), by non-rigidly transforming NIMH Macaque Template (NMT v2) into each monkey's brain MRI space. The skull-stripped monkey brain MRI in the native space then went through DRAMMS deformable registration to be spatially normalized into the NMT atlas space. This split the monkey brain MRI into a registered intensity image (i.e., the contrast channel) and the RAVENS map (i.e., the morphometry channel) similar to those in human brain MRIs. The right panel of Figure 1B shows exemplar monkey subjects for their images and the two channels we used for age estimation.

After cropping unused background in the pre-processed images, the final dimensions were 100 × 128 × 156 for monkey images in the NMT atlas space and 120 × 130 × 170 for brain images in the SRI24 atlas space. For the purpose of transfer learning, further resizing was done to linearly scale (not transforming) the human brain images to have the same size as monkey brains.

To train the deep learning model, the intensity map and RAVENS map are concatenated as one input image with two channels, with the size of 2 × 100 × 128 × 156.

Five different DL AI models were evaluated for monkey brain age estimation. They included the general 2D/3D ResNet models (He et al., 2016), which were not specifically designed for but have been previously validated for brain age estimation (He et al., 2021a, 2022). We also included three models that were specifically designed for brain age estimation: 3D simple fully convolutional network (SFCN) (Peng et al., 2021), 2D global-local transformer (GL-Transformer) (He et al., 2021a), and 3D deep relation transformer (Relation-Transformer) (He et al., 2022). For the 2D models, we consider the 3D MRIs as a stack of 2D slices with different channels.

We chose these five AI models for their representativeness. On one note, these five models represent 2D and 3D typical deep learning architectures. The 2D ResNet and GL-Transformer are 2D deep learning architectures. The 3D ResNet, SFCN, and Relation-Transformer are 3D deep learning architectures. Moreover, these five models represent both the convolutional neural network (CNN) and transformer-based algorithms. The 2D/3D ResNet and SFCN are CNN-based, whereas the GL-Transformer and Relation-Transformer are based on transformers, which are recently popular deep learning architectures based on attention mechanisms (Vaswani et al., 2017; He et al., 2022). Loosely speaking, CNN uses voxel/pixel arrays, whereas transformer models split the images into visual tokens. They are two main screams of deep learning models for image computing.

The algorithms behind these models are briefly described below. The ResNet model contains residual block with skip connection to reuse the feature feeding into the residual block. The SFCN is a short version of the VGGnet (Simonyan and Zisserman, 2014), with only five convolutional layers to extract deep features. Global-Local transformer contains a transformer (Vaswani et al., 2017) block to optimal fuse global context and local information from patches to explore more detailed brain age information on brain MRIs. Relation-Transformer (He et al., 2022) uses the attention mechanism to learn four relations, such as the cumulative, relative, maximal, and minimal relations from a pair of input images.

The models trained from scratch on monkey MRIs used the same training configurations as in human MRIs: batch size was 8, the initial learning rate of the Adam optimizer (Kingma and Ba, 2014) was 0.0001 and was reduced to half at every 50 epochs with a total of 200 training epochs.

Pre-training was done in human brain MRIs. All five models were trained with the mean absolute error (MAE) as the loss function, the same as in other human brain age estimation studies (Bashyam et al., 2020; Feng et al., 2020; He et al., 2021a,b, 2022; Peng et al., 2021). Parameters on models were optimized using the Adam optimizer (Kingma and Ba, 2014), with an initial learning rate of 0.001, which was reduced to half at every 20 epochs with a total of 80 training epochs. The batch size was set to 16. All other parameters were set to default provided by the PyTorch platform, similar to existing studies (He et al., 2021a,b, 2022).

Transfer learning was done using the pre-trained models from human brain MRIs as an initialization. The refinement was done in two stages. In Stage 1, we refined the last fully connected layer from scratch for monkey age estimation. We fixed the parameters on convolutional layers for feature extraction. This was done with 100 epochs on monkey brain MRIs with a learning rate of 0.001. After that, Stage 2 fine-tuned the parameters in all layers of the deep learning AI models with a learning rate of 0.0005, which was quartered at the 100^th epoch. All models were trained with 200 epochs. Although the age ranges of human and monkey are not the same, age matching is not necessary for transfer learning since the last fully connected layer is randomly initialized for monkey and fine-tuned in transfer learning.

Ten AI strategies (five AI models, each with or without transfer learning) were compared.

Accuracy. We used a 10-fold cross-validation to measure the accuracy. Given N monkey brain MRIs, we randomly split them into 10 equal numbers of non-overlapping samples. Each time, one fold was selected for testing, and the rest were used for training. One metric for accuracy was the mean absolute error (MAE), MAE=1N∑iN|ŷi-yi|, where ŷ is the AI-estimated age, while y is the chronological age. Smaller MAE corresponds to higher accuracy. The second accuracy metric we used was Spearman's correlation coefficient between the chronological and AI-estimated monkey brain ages. Spearman's correlation coefficient ranges from 0 (lowest accuracy) to 1 (highest accuracy). In addition to reporting the overall accuracy across the lifespan in the main text, we also computed the accuracy metrics in different age groups in Supplementary material. A monkey's lifespan capacity is approximately 0–35 years, with a rough ratio of 1:3 to human lifespan ages (Tigges et al., 1988). Therefore, we defined and compared three age groups within the current cohort, as previously defined (Simmons, 2016): young monkeys (age≤5, n = 88), middle-aged (5 <age< 20, n = 157), aged (age≥ 20, n = 45).

Stability. Each AI strategy was applied to varying sample sizes in our monkey brain MRI dataset. A strategy is more stable if the accuracy dropped the least going from using all 290 monkey brain MRIs to using only 10, 20, 30, …, and 90% of them. Training samples were randomly selected on each fold (10-fold cross-validation). For example, when sampling 20% for training, we randomly selected 20% samples from the 9-fold for training and the rest fold for testing and repeat 10 times.

Ranking by accuracy and stability. We ranked the 10 AI strategies from 1 (least preferable) to 10 (most preferable) in each of the four accuracy metrics and three stability metrics. The total rank score is a summation of those 7 scores ranging from 7 (least preferable) to 70 (most preferable). The four accuracy metrics were MAE ranked from highest (rank score 1) to the lowest (rank score 10) when using 100% of the monkey brain MRIs; Spearman's ρ ranked from the lowest (rank score 1) to the highest (rank score 10) when using 100% of the monkey brain MRIs; MAE ranked from highest (rank score 1) to the lowest (rank score 10) when using 30% of the monkey brain MRIs; and Spearman's ρ ranked from the lowest (rank score 1) to the highest (rank score 10) when using 30% of the monkey brain MRIs. The three stability metrics were the increase of MAE ranked from the highest (rank score 1) to the lowest (rank score 10) from using 100% to 30% of the monkey brain MRIs, the increase of MAE ranked from the highest (rank score 1) to the lowest (rank score 10) from using 100% to 20% of the monkey brain MRIs, and the increase of MAE ranked from the highest (rank score 1) to the lowest (rank score 10) from using 100% to 10% of the monkey brain MRIs. The AI strategies with the highest ranking scores (ideally close to 70) were recommended among the 10 strategies.

For further interpretation of the brain age models, saliency maps were generated to show the most informative voxels from MRI scans selected by the AI model for brain age estimation. In brief, the integrated gradients were used (Sundararajan et al., 2017) to visualize the relationship between input images and predictions. The baseline was the image with zero values and a sequence of images was generated by a liner interpolation between the baseline and input image, yielding a sequence of output prediction and a sequence of gradient maps on input images.

The saliency maps shown in Figure 4 were computed as the average saliency maps across all subjects in corresponding age groups (<5 and ≥5 years for monkeys and <15 and ≥15 for humans). To quantify the tissue-specific and regional contribution to brain age estimations, we applied lobular segmentations for gray and white matter regions, as well as the inclusion of the subcortical structure as a separate region. Mean saliency values normalized by whole sample minimum and maximum saliency values for each age group were quantified for tissue-based regional segmentation and plotted to visualize the trajectories of mean saliency.

Topographical saliency maps are prepared for two age groups for each channel used in AI models (intensity/contrast & RAVENS/morphometry) and captured from three different axes.

Similar procedures were carried out to generate human saliency maps and quantified saliency trajectories in matching age groups (under 15 years old and 15 or older) (Figure 4).

The benchmark task was human brain age estimation (Figure 1A, left panel). It was based on using 8,859 T1-weighted human brain MRIs 0-97 years of age. The target task was monkey brain age estimation (Figure 1A, right panel), using T1-weighted monkey brain MRI from 290 monkeys with a lifespan of 1–30 years of age. We applied minimum pre-processing to split each bias-corrected and skull-stripped T1-weighted MRI into a 3D intensity image (for contrast information) and a 3D RAVENS (regional analysis of volumes examined in normalized space) map (for morphometry information) in the human or monkey atlas spaces (Seidlitz et al., 2018; He et al., 2021a,b, 2022; Jung et al., 2021), as shown in Figure 1B. Age estimation in this study, with or without transfer learning, was based on concatenating and using these two channels of images.

The five AI models we used in the study included two 2D AI models [the classic 2D ResNet He et al., 2016, our 2D global-local transformer (GL-Transformer) He et al., 2021a] and three 3D AI models (the classic 3D ResNet He et al., 2016, a recently-developed 3D SFCN Peng et al., 2021, and our 3D Relation Transformer He et al., 2022). These five models also include three convolutional neural network models (2D/3D ResNet, 3D SFCN) and two transformer models, that are the two main branches of deep learning AI architectures.

We performed transfer learning by first fixing the parameters from the pre-trained benchmark (human) model on convolutional layers, to refine the last fully-connected layer from scratch, and then performed refinement of all the parameters during target (monkey) model training. Figure 2 shows that transferring AI models from human to monkey brain MRIs (dashed curves) led to reduced age estimation errors (y-axis) for all five AI models compared to training directly in monkey data from scratch (solid curves). The effect varied, though, by sample size in the target task (x-axis) and across models.

The effect of transfer learning on AI accuracy was more pronounced when we used only <150, especially <50 (p < 0.05 for t-Test), randomly-sampled monkey MRIs. The benefit of transfer learning became smaller or even vanished for some AI models when using more than 200 randomly-sampled monkey MRIs. This general trend is consistent with all five models we tested although the specific cutoff sample size varied for different underlying AI models, as described below.

We evaluated the performances of the AI models using mean absolute error (MAE) and Spearman's correlation coefficient (ρ) between the actual and the AI-estimated ages. A lower MAE and a higher ρ usually correspond to higher accuracy. The lowest MAE in the analysis of the 290 monkey brain MRIs occurred when using 3D ResNet with transfer learning (MAE=1.83 ± 0.39 years). The highest Pearson correlation ρ occurred when using the 3D SFCN with transfer learning (ρ = 0.921).

Two 3D AI models (3D SFCN and 3D relation transformer) benefited more from transfer learning than the two 2D AI models (2D ResNet and 2D GL-Transformer). In general, 3D AI models often required larger sample sizes than 2D models (Schulz et al., 2020).

The AI model that benefited the most was the model with the largest errors (3D SFCN). AI models that had a relatively smaller error (<4 years), even with just 30 randomly-sampled monkey brain MRIs, benefited less from transfer learning.

Of special note is the 3D relation transformer, which, given N input images, used N² pairs of images as training samples. The 3D relation transformer estimated the relationship (sum, difference, bigger/smaller than) on every pair of input images (He et al., 2022). Therefore, transfer learning reduced errors compared to without transfer learning for the 3D relation transformer, with 220 or fewer input images (<48,400 input image pairs). The MAE of transfer learning became larger than the MAE without transfer learning with more than 220 input monkey brain MRIs (more than 48,400 input MRI pairs).

In general, the comparison across the underlying AI models showed that those AI models that needed more input MRIs (e.g., 3D SFCN and 3D ResNet) would benefit more from transfer learning.

We next ranked the 10 strategies (five AI models; with or without transfer learning) based on their accuracy and stability (Figure 3). The rank score ranged from 1 (least recommended) to 10 (most recommended) in four accuracy metrics and three stability metrics (with each row being a metric). Accuracy was measured as the mean absolute errors (MAE, smaller is better) and Spearman's correlation (higher is better) at 30% and 100% of the samples between AI-estimated and actual chronological monkey ages. Stability was measured by the increase of MAE (smaller is better) from when using 100% of the samples to when using 10%, 20%, and 30% of the samples. Overall, the 2D GL-Transformer with transfer learning and the 3D ResNet with transfer learning received the highest total rank scores, both scored 56 out of 70. They were followed by 2D ResNet with transfer learning, with a total rank score of 51. These recommended choices were all with transfer learning.

After examining the accuracy and stability of different AI models, we visualized the key MRI regions and features (from the contrast and morphometry channels) that contributed the most to age estimation in the best-performing model, 3D ResNet model with transfer learning, which had the highest rank scores in Figure 3 and the lowest MAE when all 290 MRIs were used in cross-validation. We generated saliency maps, which are topographical representations of feature importance, showing the most informative voxels selected by the AI model for brain age estimation (Figure 4). Three main observations are noted when comparing saliency maps generated from the pre-trained human model: