ReBA-Pred-Net: Weakly-Supervised Regional Brain Age Prediction on MRI

Brain age has become a prominent biomarker of brain health. Yet most prior work targets whole brain age (WBA), a coarse paradigm that struggles to support tasks such as disease characterization and research on development and aging patterns, because relevant changes are typically region-selective rather than brain-wide. Therefore, robust regional brain age (ReBA) estimation is critical, yet a widely generalizable model has yet to be established. In this paper, we propose the Regional Brain Age Prediction Network (ReBA-Pred-Net), a Teacher-Student framework designed for fine-grained brain age estimation. The Teacher produces soft ReBA to guide the Student to yield reliable ReBA estimates with a clinical-prior consistency constraint (regions within the same function should change similarly). For rigorous evaluation, we introduce two indirect metrics: Healthy Control Similarity (HCS), which assesses statistical consistency by testing whether regional brain-age-gap (ReBA minus chronological age) distributions align between training and unseen HC; and Neuro Disease Correlation (NDC), which assesses factual consistency by checking whether clinically confirmed patients show elevated brain-age-gap in disease-associated regions. Experiments across multiple backbones demonstrate the statistical and factual validity of our method.

💡 Research Summary

The paper addresses a critical gap in brain‑age research: while whole‑brain age (WBA) prediction has become a mature field, regional brain‑age (ReBA) estimation remains under‑developed due to the lack of region‑level ground‑truth labels. To overcome this, the authors propose ReBA‑Pred‑Net, a novel teacher‑student framework that can learn fine‑grained brain‑age maps from raw 3D T1‑weighted MRI using only the subject’s chronological age as supervision.

Teacher Module
First, a conventional WBA regressor (the “Teacher”) is trained on healthy controls (HC) to predict whole‑brain age and then frozen. The brain is parcellated with a standard functional atlas (e.g., Harvard‑Oxford). Each region is isolated, fed through the frozen Teacher, and an initial regional age estimate is obtained. To prevent the “mode‑collapse” problem—where every region would simply inherit the same whole‑brain age—the authors introduce a regional correction signal. By occluding a region with low‑amplitude noise and measuring the change in the Teacher’s whole‑brain prediction, they quantify that region’s contribution to the global age estimate. This delta is added to the initial regional estimate, yielding a soft ReBA that reflects both the Teacher’s knowledge and the region’s marginal influence.

Student Module
The Student shares the same backbone as the Teacher but augments each region with a learnable prompt embedding. These prompts modulate the backbone features via Feature‑wise Linear Modulation (FiLM), after which a lightweight adapter produces the final regional age prediction. Training the Student involves two complementary losses: (1) a distillation loss that forces the Student’s output to match the Teacher’s soft ReBA, and (2) a functional‑consistency loss that encourages regions belonging to the same functional system (e.g., visual, motor) to have similar age trajectories. This consistency term mitigates over‑averaging and preserves biologically plausible spatial gradients.

Evaluation without Region‑Level Ground Truth
Standard regression metrics such as MAE or MSE cannot be applied because true regional ages are unavailable. The authors therefore devise two indirect but meaningful evaluation metrics:

1. Healthy Control Similarity (HCS) – For an unseen set of healthy controls, the distribution of the Regional Brain Age Gap (ΔReBA = ReBA – chronological age) is compared to that of the training HC cohort. Statistical alignment indicates that the model generalizes without distributional drift and is well‑calibrated on healthy brains.

2. Neuro Disease Correlation (NDC) – Using clinically established disease priors (e.g., motor‑related regions for Parkinson’s disease, memory‑related regions for Alzheimer’s disease), ΔReBA is computed for patients and compared against HC and other disease groups. Significantly larger gaps in the disease‑relevant regions demonstrate factual consistency: the model correctly identifies that those regions appear “older” in the disease population.

Experimental Findings
The authors evaluate ReBA‑Pred‑Net with several backbones, including 3‑D CNNs and Vision Transformers. Across all architectures, HCS scores improve relative to baseline whole‑brain models, confirming statistical consistency. In NDC experiments, patients with Parkinson’s disease show elevated ΔReBA in motor cortices, while Alzheimer’s patients exhibit higher gaps in hippocampal and temporal regions; both effects are statistically significant (p < 0.01). Ablation studies reveal that removing the functional‑consistency loss leads to a collapse of regional age variance and a drop in NDC performance, underscoring the importance of the proposed constraints.

Contributions and Impact

1. Methodological Innovation – A teacher‑student pipeline that converts weak whole‑brain supervision into reliable region‑level soft labels, augmented by an occlusion‑based correction signal.
2. New Evaluation Paradigm – Introduction of HCS and NDC as complementary metrics that assess statistical calibration and factual disease relevance without requiring region‑wise ground truth.
3. Biological Plausibility – The functional‑consistency regularizer preserves known spatial gradients of brain development and aging, making the predictions interpretable for neuroscientists and clinicians.

Future Directions
Potential extensions include integrating multimodal imaging (e.g., PET, diffusion MRI), learning subject‑specific functional atlases, and deploying the model in longitudinal clinical studies to track disease progression or treatment response. Overall, ReBA‑Pred‑Net offers a practical and theoretically sound solution for regional brain‑age estimation, paving the way for more nuanced biomarkers of brain health.