Model and predict age and sex in healthy subjects using brain white matter features: A deep learning approach

Hao He, Fan Zhang, Steve Pieper, Nikos Makris, Yogesh Rathi, William M. Wells Iii, Lauren J. O'Donnell

Abstract

The human brain's white matter (WM) structure is of immense interest to the scientific community. Diffusion MRI gives a powerful tool to describe the brain WM structure noninvasively. To potentially enable monitoring of age-related changes and investigation of sex-related brain structure differences on the mapping between the brain connectome and healthy subjects' age and sex, we extract fiber-cluster-based diffusion features and predict sex and age with a novel ensembled neural network classifier. We conduct experiments on the Human Connectome Project (HCP) young adult dataset and show that our model achieves 94.82% accuracy in sex prediction and 2.51 years MAE in age prediction. We also show that the fractional anisotropy (FA) is the most predictive of sex, while the number of fibers is the most predictive of age and the combination of different features can improve the model performance.

Introduction

The human brain's white matter (WM) is a complex network of fibers that undergoes significant changes during aging and exhibits distinct differences between sexes. While Diffusion MRI (dMRI) allows for the non-invasive mapping of these structures through tractography, translating this high-dimensional, often noisy data into accurate demographic predictions remains a significant challenge.

Researchers face three primary hurdles in this domain: identifying which specific diffusion measurements (e.g., water flow directionality) signal age or sex, managing the intense preprocessing required for dMRI data, and developing models sophisticated enough to handle the high dimensionality of brain networks without overfitting.

Methodology

This study proposes a multi-step pipeline to transform raw neuroimaging data into predictive insights. The process begins with Tractography and Parcellation, using the Unscented Kalman Filter (UKF) to trace brain fibers, which are then grouped into 800 anatomically curated clusters.

From these clusters, 14 distinct features are extracted, including Fractional Anisotropy (FA), Mean Diffusivity (MD), and total fiber count. The researchers evaluated two primary Convolutional Neural Network (CNN) architectures to process these features:

  • 1D-CNN: Processes the 1,516 cluster measurements as a continuous vector.
  • 2D-CNN: Organizes clusters into a 3x800 matrix based on anatomical location (Left, Right, and Commissural).
The proposed 1D-CNN (a) and 2D-CNN (b) architectures for processing white matter cluster features.

To enhance stability and accuracy, an Ensemble Learning approach was adopted. Rather than relying on a single feature, multiple "weak" 1D-CNNs were trained on various diffusion measures and their outputs were combined using a bagging technique.

Results and Discussion

The study utilized the Human Connectome Project (HCP) Young Adult dataset, comprising 964 healthy subjects aged 22–37. Despite the narrow age range, the model demonstrated high precision.

Feature Predictive Power

Analysis revealed that different diffusion measures serve as better predictors for different demographics. Fractional Anisotropy (FA) was the most predictive feature for sex, suggesting that sex differences are primarily reflected in the microstructure and density of tracts. Conversely, the total number of fibers was the strongest predictor for age, indicating that early adult aging is most visible in the quantity of detectable fibers.

Comparison of model performance across different diffusion features for sex prediction (accuracy) and age prediction (MAE).

Model Performance

  • Sex Prediction: The ensemble 1D-CNN achieved a state-of-the-art 94.82% accuracy.
  • Age Prediction: The model achieved a Mean Absolute Error (MAE) of 2.51 years.

Interestingly, the 1D-CNN outperformed the 2D-CNN. This suggests that the spatial relationships imposed by a 2D grid may be less beneficial than the direct feature-to-feature learning of a 1D architecture when dealing with relatively independent brain clusters. Furthermore, the study found that the mean value within a cluster was significantly more predictive than extreme (min/max) values, which are often more susceptible to noise.

Conclusion

This deep learning framework provides a robust method for predicting age and sex from brain white matter features. While the results are promising, future work is needed to validate the model on "lifespan" datasets (including children and the elderly) and across different imaging sites to ensure generalizability. Additionally, refining the optimal subset of features remains an open area for research to further improve efficiency and accuracy.

Summary updated

Abstract

The human brain's white matter (WM) structure is of immense interest to the scientific community. Diffusion MRI gives a powerful tool to describe the brain WM structure noninvasively. To potentially enable monitoring of age-related changes and investigation of sex-related brain structure differences on the mapping between the brain connectome and healthy subjects' age and sex, we extract fiber-cluster-based diffusion features and predict sex and age with a novel ensembled neural network classifier. We conduct experiments on the Human Connectome Project (HCP) young adult dataset and show that our model achieves 94.82% accuracy in sex prediction and 2.51 years MAE in age prediction. We also show that the fractional anisotropy (FA) is the most predictive of sex, while the number of fibers is the most predictive of age and the combination of different features can improve the model performance.
BibTeX
@inproceedings{he2022model,
  author = {Hao He and Fan Zhang and Steve Pieper and Nikos Makris and Yogesh Rathi and William M. Wells III and Lauren J. O'Donnell},
  title = {{Model and predict age and sex in healthy subjects using brain white matter features: A deep learning approach}},
  booktitle = {ISBI},
  year = {2022},
  url = {https://openreview.net/forum?id=C5YMxRzSWJ},
}
Figure