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

TL;DR

ReBA-Pred-Net tackles the challenge of estimating regional brain age (ReBA) under weak supervision by converting whole-brain age signals into region-wise soft targets via a frozen Teacher and training a region-aware Student with FiLM prompts and a functional-consistency constraint. Two indirect evaluation metrics, Healthy Control Similarity ($HCS$) and Neuro Disease Correlation ($NDC$), provide complementary validation in the absence of region-level ground truth. Across multiple backbones and large HC datasets, the method achieves stable training, interpretable regional aging patterns, and disease-aligned findings (e.g., PD regions showing elevated ReBA). The approach offers a scalable, feature-free pathway to regional brain aging analytics with practical clinical relevance, while acknowledging limitations in data quality, priors, and cross-site generalization that warrant further work.

Abstract

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.

Figures (12)

• Figure 1: (a) Motivation. The coarse WBA cannot adequately support disease-level analyses or studies of aging. Parcellating the brain with established functional atlases and modeling ReBA has the potential to solve these limitations and to strengthen brain-health research. (b) Challenges. ReBA lacks region-level ground truth. In training, HC chronological age is the only available weak signal, risking mode collapsing and obscuring true regional differences; in evaluation, the absence of regional truth prevents region-wise supervision, making standard metrics (e.g., MAE/MSE) inapplicable. (c) Solutions. We propose ReBA-Pred-Net with Teacher-Student framework to address the training problem, and introduce two evaluation metrics, HCS and NDC, to assess statistical consistency and factual consistency.
• Figure 2: The flowchart of Regional Brain Age Prediction Network (ReBA-Pred-Net). Teacher module consists of three decoupled steps: (i) Feed the whole-brain MRI to the Teacher to predict WBA and train it against chronological age; (ii) Parcellate the brain using an atlas; extract each region and pass it through the frozen Teacher to obtain initial ReBA; (iii) Apply the corrections in Eqs. (\ref{['eqa: regional correction signal']}), (\ref{['eqa: teacher_final_regional_age']}) to each region to produce the final soft ReBA. Student module comprises three sequential steps: (i) Feed each region to the Student Main (shared with the Teacher) to obtain a regional embedding; (ii) Add a learnable prompt per region, generate modulated parameters with a tunable Student FiLM, and fuse them with the regional embedding; (iii) Pass the fused feature through a lightweight adapter to produce the final fine-grained ReBA. We optimize the Student with a distillation loss (to match the Teacher’s soft ReBA) and a functional-consistency loss (enforcing that regions within the same functional network change similarly). Notably, all training is conducted exclusively on HC.
• Figure 3: Healthy Control Similarity (HCS) across brain regions (see Sec. \ref{['sec: Supplemented Experiments']} for more regions' results). Results are based on the 3D DenseNet backbone lee2022deep. Blue bars denote the $\Delta$ReBA distributions predicted by our model on training HC; red bars denote the corresponding predictions on test HC. Statistical similarity is computed via Eq. (\ref{['eqa: region-wise consistency score']}). Higher is better. The results show consistently high statistical similarity at the regional level, with an overall HCS reaching 73%, which supports the effectiveness of our method.
• Figure 4: Neuro-Disease Correlation (NDC) for Parkinson’s disease (PD). We focus on regions 3, 4, 7, 26, all of which show clinical evidence of accelerated aging in PD. Accordingly, the $\Delta$ReBA in these regions is expected to exceed those observed in healthy controls (HC) and Alzheimer’s disease (AD). We compute the NDC and compare it across PD, HC, and AD. As anticipated, PD exhibits a markedly higher NDC than both HC and AD, providing indirect support for the accuracy of the proposed ReBA prediction.
• Figure 5: Failure case visualization. (a) and (c) are failures: (a) PD with low NDC resembles the HC in (b); (c) HC with high NDC resembles the PD in (d). Morphometric similarity in PD-related regions likely drives the unexpected NDC.