Estimating Brain Age with Global and Local Dependencies

Yanwu Yang, Xutao Guo, Zhikai Chang, Chenfei Ye, Yang Xiang, Haiyan Lv, Ting Ma

Abstract

Brain age is a critical phenotype linked to cognitive performance and neurological health. Accurate brain age prediction is essential for utilizing the predicted brain-age difference (PAD) as a reliable biomarker. However, capturing the complexities of brain aging is challenging for models relying solely on local feature engineering or standard convolutional operations. This paper introduces the Global and Local Dependency Network (GLDN), a novel architecture designed to interpret brain age by integrating global and local representations. The network utilizes a Successive Permuted Transformer (SPT) to capture global attentive interactions and convolution blocks for local feature extraction. The SPT enhances computational efficiency by indirectly locating 3D spatial information through the continuous encoding of 2D slices from multiple views. Evaluated on a large cohort of 22,645 subjects (ages 14–97), the GLDN achieved a Mean Absolute Error (MAE) of 2.855 on the validation set and 2.911 on an independent test set, outperforming several state-of-the-art deep learning methods.

Introduction

In modern neurology, chronological age does not always align with biological brain health. The concept of "brain age"—the age a brain appears to be based on structural neuroimaging—has emerged as a potent biomarker. The gap between chronological age and predicted brain age, known as the Predicted Age Difference (PAD), serves as an indicator of neurobiological status. A positive PAD, where the brain appears older than the individual's actual age, is frequently associated with cognitive decline, traumatic brain injury, and neurodegenerative conditions such as Alzheimer’s disease.

Estimating brain age is a fine-grained regression task that requires a model to understand both minute local changes in tissue density and large-scale structural relationships across the entire brain. Traditional Convolutional Neural Networks (CNNs) excel at extracting local textures but often fail to model long-range dependencies across the 3D volume. Conversely, Vision Transformers (ViTs) are adept at capturing global context but face significant computational hurdles when processing high-resolution 3D medical images. The Global and Local Dependency Network (GLDN) was developed to bridge this gap, combining the strengths of both architectures.

Methodology

The Global and Local Dependency Network (GLDN)

The GLDN architecture is built upon three primary components designed to balance local detail with global structural context:

  1. Local Learning Block (LLB): This module utilizes 3D CNN layers to encode dense local features. By leveraging the inductive bias of "locality," the LLB ensures the model remains sensitive to nearby voxel relationships, which are crucial for identifying subtle structural changes.
  2. Global Learning Block (GLB) and Successive Permuted Transformer (SPT): To overcome the memory and computational costs of 3D self-attention, the authors introduced the SPT. This module permutes the 3D brain volume into 2D slices from three orthogonal perspectives: Sagittal, Axial, and Coronal. By processing these views sequentially, the Transformer captures 3D spatial relationships through a series of efficient 2D operations.
  3. Fusion Blocks: These modules aggregate outputs from the LLB and GLB at multiple stages of the network. This multi-scale fusion ensures that the final prediction is informed by both fine-grained local textures and broad global dependencies.

Label Distribution Learning (LDL)

Rather than treating brain age as a single scalar value for regression, the GLDN employs Label Distribution Learning. This approach treats age as a probability distribution, acknowledging the inherent ambiguity in biological aging. By predicting a distribution rather than a single point, the model better accounts for the fact that a healthy brain may share characteristics with slightly younger or older cohorts.

Experimental Setup and Datasets

The study utilized an extensive dataset of 22,645 healthy subjects, one of the largest cohorts used in brain age research. The data was aggregated from five major neuroimaging repositories:

  • UK Biobank
  • ADNI (Alzheimer’s Disease Neuroimaging Initiative)
  • IXI Database
  • OASIS (Open Access Series of Imaging Studies)
  • 1000 Functional Connectomes Project (1000-FCP)

The subjects spanned a wide age range from 14 to 97 years, providing a comprehensive representation of the human brain across the lifespan.

Results and Discussion

The GLDN demonstrated superior performance compared to existing state-of-the-art models. In a comparative analysis, the GLDN achieved a Mean Absolute Error (MAE) of 2.911 on the independent test set and a Pearson Correlation Coefficient (PCC) of 0.879.

Comparative Performance

  • GLDN (Ours): 2.911 MAE
  • ResNet-34: 3.334 MAE
  • Vision Transformer (ViT): 3.536 MAE
  • SFCN (Standard Brain-Age CNN): 3.068 MAE

Ablation studies confirmed the necessity of the hybrid approach. Removing the CNN component (the "Ours w/o CNN" configuration) led to a significant drop in accuracy, highlighting that global context alone cannot compensate for the loss of local structural information. Furthermore, the SPT's permutation strategy proved to be an efficient alternative to standard 3D Transformers, maintaining high accuracy while significantly reducing the computational footprint.

Conclusion

The GLDN represents a significant advancement in brain age estimation by successfully integrating global and local dependencies. By utilizing the Successive Permuted Transformer and Label Distribution Learning, the model achieves high precision across a massive, diverse dataset. Future work will focus on applying this model to clinical populations to determine if its increased accuracy translates into higher sensitivity for detecting early-stage neurodegenerative diseases and other brain-related pathologies.

Summary updated

Abstract

The brain age has been proven to be a phenotype of relevance to cognitive performance and brain disease. Achieving accurate brain age prediction is an essential prerequisite for optimizing the predicted brain-age difference as a biomarker. As a comprehensive biological characteristic, the brain age is hard to be exploited accurately with models using feature engineering and local processing such as local convolution and recurrent operations that process one local neighborhood at a time. Instead, Vision Transformers learn global attentive interaction of patch tokens, introducing less inductive bias and modeling long-range dependencies. In terms of this, we proposed a novel network for learning brain age interpreting with global and local dependencies, where the corresponding representations are captured by Successive Permuted Transformer (SPT) and convolution blocks. The SPT brings computation efficiency and locates the 3D spatial information indirectly via continuously encoding 2D slices from different views. Finally, we collect a large cohort of 22645 subjects with ages ranging from 14 to 97 and our network performed the best among a series of deep learning methods, yielding a mean absolute error (MAE) of 2.855 in validation set, and 2.911 in an independent test set.
BibTeX
@inproceedings{yang2022estimating,
  author = {Yanwu Yang and Xutao Guo and Zhikai Chang and Chenfei Ye and Yang Xiang and Haiyan Lv and Ting Ma},
  title = {{Estimating Brain Age with Global and Local Dependencies}},
  booktitle = {ICIP},
  year = {2022},
  doi = {10.1109/icip46576.2022.9897454},
}
Figure