Brain age prediction based on resting-state functional connectivity patterns using convolutional neural networks

Hongming Li, Theodore D. Satterthwaite, Yong Fan

This paper, authored by researchers from the University of Pennsylvania, presents a deep learning framework for predicting "brain age"—a biological proxy for neurodevelopment—using high-resolution functional connectivity data.

TL;DR

Researchers have developed a deep learning model that predicts a person's age by analyzing fine-grained, voxel-wise functional connectivity maps from resting-state fMRI. By using Convolutional Neural Networks (CNNs) to capture complex spatial patterns across 56 different functional networks, the model achieves superior accuracy compared to traditional methods that rely on coarse regional averages.

What Problem Does This Paper Address?

"Brain age" is the predicted age of an individual based on neuroimaging data. When a person’s "brain age" differs significantly from their chronological age, it can serve as a biomarker for atypical development or neuropsychiatric disorders.

While structural MRI (looking at brain shape and volume) has been used for this in the past, functional connectivity (FC)—which measures how different brain regions "talk" to each other—offers a dynamic view of brain maturation. However, most existing FC studies use "coarse-grained" measures, averaging signals within large Regions of Interest (ROIs). This paper argues that these averages discard "fine-grained" information. The authors set out to prove that voxel-wise (pixel-level) connectivity patterns across the whole brain contain vital information that can significantly improve age prediction.

Methods

The authors propose a multi-step pipeline that combines traditional neuroimaging analysis with modern deep learning:

  1. ICN Identification: They used collaborative non-negative matrix factorization (NMF) to identify 56 subject-specific Intrinsic Connectivity Networks (ICNs). This ensures that while the networks are consistent across the group, they respect the unique functional anatomy of each individual.
  2. Dense FC Mapping: For each of the 56 ICNs, they calculated the Pearson correlation between that network's time course and every other voxel in the brain. This resulted in 56 "dense" connectivity maps per subject.
  3. 4D CNN Architecture: These 56 maps were stacked as multi-channel inputs (similar to how a color image has Red, Green, and Blue channels) into a deep CNN.
  4. Architecture: The network consisted of one initial convolutional layer, followed by three Residual Blocks (ResBlocks), a fully connected layer, and a final output layer for age regression.
  5. Optimization: The model was trained using Stochastic Gradient Descent (SGD) with Euclidean loss to minimize the difference between predicted and actual age.

Datasets

The study utilized a large-scale dataset from the Philadelphia Neurodevelopmental Cohort (PNC):

  • Subjects: 983 individuals.
  • Age Range: 8 to 22 years old (capturing the critical window of adolescent brain maturation).
  • Modality: Resting-state functional MRI (rsfMRI).

Key Insights and Ideas

  • Fine-Grained Info is King: The move from ROI-to-ROI connectivity to voxel-wise connectivity was the primary driver of performance. The CNN was able to "see" spatial nuances that linear models and coarse averages miss.
  • Hierarchical Feature Learning: By using ResBlocks, the model could learn hierarchical patterns—starting from local connectivity "textures" and moving up to global network interactions.
  • Identifying "Clock" Networks: Through a PCA-based sensitivity analysis, the authors identified which networks were most influential for the "clock." The cingulo-opercular, default mode, and attention networks emerged as the most sensitive predictors, aligning with known neurodevelopmental literature regarding the maturation of executive function and self-referential thought during adolescence.

Key Results

The CNN model outperformed several baselines, including standard Lasso regression and BrainNetCNN (a specialized architecture for network graphs).

Model Features Correlation ($r$) MAE (Years)
Lasso Inter-ICN FC (Coarse) 0.417 2.45
E2Enet-sml Inter-ICN FC (Coarse) 0.361 3.02
Lasso Whole Brain FC (Fine) 0.530 2.32
CNNs (Ours) Whole Brain FC (Fine) 0.614 2.15

The Mean Absolute Error (MAE) of 2.15 years represents a significant improvement in precision for a functional-only model in this age range. Furthermore, t-SNE visualizations showed that the features learned by the CNN clustered subjects by age much more cleanly than the raw connectivity data.

Limitations and Future Work

  • Computational Intensity: Processing whole-brain voxel-wise maps is computationally expensive. The authors had to down-sample the maps to $4 \times 4 \times 4\text{mm}$ resolution to keep the model tractable.
  • Single Modality: This study focused purely on functional data. The authors note that integrating this with structural MRI (multimodal fusion) would likely yield even higher predictive power.
  • Interpretability: While the sensitivity analysis helps, deep CNNs remain somewhat of a "black box" compared to simple linear regression, making it harder to pinpoint exactly which anatomical connections are changing with age.

Summary updated

Abstract

Brain age prediction based on neuroimaging data could help characterize both the typical brain development and neuropsychiatric disorders. Pattern recognition models built upon functional connectivity (FC) measures derived from resting state fMRI (rsfMRI) data have been successfully used to predict the brain age. However, most existing studies focus on coarse-grained FC measures between brain regions or intrinsic connectivity networks (ICNs), which may sacrifice fine-grained FC information of the rsfMRI data. Whole brain voxel-wise FC measures could provide fine-grained FC information of the brain and may improve the prediction performance. In this study, we develop a deep learning method to use convolutional neural networks (CNNs) to learn informative features from the fine-grained whole brain FC measures for the brain age prediction. Experimental results on a large dataset of resting-state fMRI demonstrate that the deep learning model with fine-grained FC measures could better predict the brain age.
BibTeX
@inproceedings{li2018brain,
  author = {Hongming Li and Theodore D. Satterthwaite and Yong Fan},
  title = {{Brain age prediction based on resting-state functional connectivity patterns using convolutional neural networks}},
  booktitle = {ISBI},
  year = {2018},
  url = {https://openreview.net/forum?id=ym4D0HFja2Q},
}

Loading related content...

Figure