Specialized White and Gray Matter Brain Age Models Using Conditional Variational Autoencoder

1926 

Abstract Submission 

Jinghang Li¹, Chang-Le Chen¹, Linghai Wang¹, Shaolin Yang¹, Howard Aizenstein¹, Minjie Wu¹

¹University of Pittsburgh, Pittsburgh, PA

Jinghang Li  
University of Pittsburgh
Pittsburgh, PA

Chang-Le Chen  
University of Pittsburgh
Pittsburgh, PA

Linghai Wang  
University of Pittsburgh
Pittsburgh, PA

Shaolin Yang, PhD  
University of Pittsburgh
Pittsburgh, PA

Howard Aizenstein  
University of Pittsburgh
Pittsburgh, PA

Minjie Wu  
University of Pittsburgh
Pittsburgh, PA

Brain age (BA) prediction models have emerged as valuable tools for understanding individual differences in brain aging trajectory. These models, typically constructed using machine learning techniques, aim to estimate the overall brain health by predicting brain age based on structural MRI data (Lee, Burkett et al. 2022). Using these BA models, women are found to be 1.5-3.5 years younger than men (Lee, Burkett et al. 2022). This obviously does not align with the epidemiologic findings that women are disproportionately affected by dementia (Zhu, Montagne et al. 2021). This suggests that existing BA models are not sufficient to characterize brain aging. Therefore, in this study, we designed and implemented two distinct deep learning-based brain age models with one focusing on the cortical gray matter (GM) while the other on the cerebral white matter (WM). We hypothesize that our WM-BA model can better capture small vessel disease (SVD) related brain aging than GM-BA models.

Both the gray matter and the white matter brain age prediction models take in 3D volumetric input images. While gray matter model takes in cortical thickness map (Rebsamen, Rummel et al. 2020), the white matter model takes in myelin ratio map (Hannoun, Kocevar et al. 2022). We constructed age conditioned variational autoencoder with 3D convolution blocks for the encoder-decoder component and multilayer perceptron (MLP) for the brain age prediction component (see figure 1 for detailed model schematics). We trained our models on the HCP healthy aging dataset (female, n=396; male, n=316) (Bookheimer, Salat et al. 2019) with an 80-20 training-testing ratio to mitigate overfitting. The trained models' adaptability was tested by conducting model inference on the Cam-CAN healthy aging dataset (Shafto, Tyler et al. 2014), and the credibility of BA predictions was validated through examination of the models' encoded latent space. Additionally, we explored correlations between brain age gaps and white matter lesion burdens (white matter lesion volume/intracranial volume). Both the white matter lesions and the brain segmentations were extracted from FreeSurfer (Fischl 2012). All brain age predictions have been bias corrected via fitted linear regression models (Beheshti, Nugent et al. 2019).

Figure 1 illustrates the generated cortical thickness map and myelin ratio map, along with the model's brain age predictions and reconstructed structural outputs. Figure 2 visualizes the model's latent vectors from the Cam-CAN dataset and revealed effective disentanglement on the age feature, enhancing MLP's predictability (Zhao, Adeli et al. 2019) (figure 2.a., 2.c.). Further, the linear regression models between brain age gap and the white matter lesion burdens illustrates that WM-BA gap correlates with white matter lesions (p=0.000001, r=0.25) better than GM-BA gap (p=0.0003, r=0.19), suggesting compared to gray matter brain age model, white matter brain age model can better capture cerebrovascular aging (figure 2.b., 2.d.). Lastly, our brief exploratory analysis showed that women can be more resilient in gray matter decay than men do in healthy aging, suggesting gray matter brain age models fail to capture SVD-related brain aging in women (figure 2.b., 2.d.).

In summary, this work demonstrates the potential avenue for developing specialized brain aging models that facilitate better understanding in brain aging trajectories across different regions as well as across sex.

Aging ²

Methods Development ¹

Segmentation and Parcellation

Anatomical MRI

Aging

Cortical Layers

Machine Learning

Myelin

White Matter