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

TL;DR

Researchers have developed a deep learning framework that predicts a person’s sex and age by analyzing the "wiring" of the brain’s white matter. By parcellating brain fibers into anatomical clusters and using an ensemble of convolutional neural networks (CNNs), the model achieved nearly 95% accuracy in sex prediction and a mean error of just 2.5 years for age.

What Problem Does This Paper Address?

The human brain's white matter (WM) is a complex network of fibers that changes as we age and differs between sexes. While Diffusion MRI (dMRI) allows us to map these structures non-invasively through "tractography," turning this high-dimensional, noisy data into accurate demographic predictions is difficult.

Previous studies often relied on simple global averages or graph-based models that might miss fine-grained anatomical details. The authors identify three main hurdles:

  1. Feature Selection: Which specific diffusion measurements (like water flow speed or directionality) actually signal age or sex?
  2. Extraction Complexity: dMRI data requires intense preprocessing, and different methods yield different results.
  3. High Dimensionality: Brain networks are massive; models need to be sophisticated enough to handle the complexity without overfitting.

Methods

The authors proposed a multi-step pipeline that moves from raw images to deep learning:

  • Tractography & Parcellation: They used a technique called the Unscented Kalman Filter (UKF) to trace brain fibers. These fibers were then automatically grouped into 800 anatomically curated "clusters" (e.g., specific bundles in the left hemisphere, right hemisphere, or connecting the two).
  • Feature Engineering: For each cluster, they extracted 14 different features, including Fractional Anisotropy (FA - a measure of how "directional" water diffusion is), Mean Diffusivity (MD), and the total number of fibers.
  • CNN Architectures: They compared two models:
  • 1D-CNN: Treats the 1,516 cluster measurements as a long vector.
  • 2D-CNN: Arranges clusters into a 3x800 matrix based on anatomical location (Right, Left, and Commissural).
  • Ensemble Learning: Instead of picking one "best" feature, they trained multiple "weak" 1D-CNNs on different diffusion measures and combined their outputs. This "bagging" approach improves stability and accuracy.

Datasets

The study utilized the Human Connectome Project (HCP) Young Adult dataset.

  • Sample Size: 964 healthy subjects (443 male, 521 female).
  • Age Range: 22–37 years old.
  • Significance: This is a high-quality, gold-standard dataset in neuroimaging, though its narrow age range provides a specific challenge for age regression.

Key Insights and Ideas

  • FA vs. MD: Fractional Anisotropy (FA) proved to be the most predictive feature for sex, while the total number of fibers was the strongest predictor for age. This suggests that sex differences are more reflected in the microstructure (integrity/density) of tracts, while early adult aging is more visible in the quantity of detectable fibers.
  • The Power of the Mean: When looking at diffusion measures, the mean value within a cluster was significantly more predictive than the maximum or minimum values. Extreme values are likely noisier and less representative of the underlying anatomy.
  • 1D beats 2D: Surprisingly, the 1D-CNN outperformed the 2D-CNN. The authors suggest that because brain clusters are relatively independent, the "spatial" relationship forced by a 2D grid might not be as beneficial as the direct feature-to-feature learning in a 1D architecture.
  • Ensembles Matter: Combining different types of diffusion data (like FA and fiber counts) consistently outperformed models using a single feature type.

Key Results

  • Sex Prediction: The ensemble 1D-CNN achieved 94.82% accuracy, outperforming standard benchmarks like LeNet.
  • Age Prediction: The model achieved a Mean Absolute Error (MAE) of 2.51 years. Given that the subjects were all within a tight 15-year window (22-37), this level of precision is quite high.
  • State of the Art: The authors' method represents the state-of-the-art (SOTA) for connectome-based (fiber-based) prediction, though it slightly trails some models that use raw T1-weighted structural images (which contain different types of anatomical information).

Limitations and Future Work

  • Narrow Age Range: The HCP Young Adult dataset only covers ages 22–37. The authors acknowledge that the model needs to be tested on "lifespan" datasets (children to elderly) to see if these white matter features remain the best predictors across all stages of life.
  • Single Dataset: The study relied on one high-quality dataset. Future work should involve "cross-site" validation to ensure the model isn't just learning the specific quirks of the HCP scanners.
  • Feature Pruning: While the ensemble method is powerful, the authors noted that simply adding more features doesn't always help; identifying the optimal subset of features remains an open area for refinement.

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 non-invasively. 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},
}

Loading related content...

Figure