# From Infancy to Aging: Precise Brain Age Estimation via Hybrid CoTResNet3D and CrossViT Models on T1-Weighted Imaging

## TL;DR

This paper introduces a new hybrid AI model that accurately estimates brain age from MRI scans across the entire lifespan, from infancy to aging.

## Contribution

A novel hybrid ResNet-CrossViT model is proposed for precise brain age estimation with strong generalization and reliability.

## Key findings

- The model achieved a mean absolute error of 2.72 years on the internal test set.
- It showed strong cross-center generalization with a mean absolute error of 4.19 years.
- The model demonstrated excellent test–retest reliability with an ICC of 0.994.

## Abstract

Accurate estimation of brain age from structural magnetic resonance imaging (MRI) serves as a vital biomarker for quantifying individual neurobiological aging and identifying risks for neurological disorders. However, developing robust models that generalize across the entire lifespan (from infancy to aging) remains challenging due to heterogeneous maturation/degeneration patterns, limited cross-center generalizability, and insufficient temporal reliability evaluation. To address these limitations, we curated a large-scale, multi-center T1-weighted MRI dataset across 27 public cohorts. Of these, 22,271 scans from 17 cohorts (aged 0–96 years) formed the primary foundation for model development, complemented by 10 additional cohorts utilized for independent multi-center evaluation and robustness testing. We propose ResNet-CrossViT, a novel hybrid architecture that synergistically combines a 3D Contextual Transformer-ResNet (CoTResNet3D) backbone for enriched local feature extraction and a CrossVision Transformer (CrossViT) module for cross-scale global dependency modeling. The model was rigorously evaluated on an internal test set, an unseen external dataset for cross-center validation, a longitudinal dataset for assessing temporal consistency, and a test–retest dataset for measuring reproducibility. On the internal test set, ResNet-CrossViT achieved a mean absolute error (MAE) of 2.72 years and a maximal MAE (mMAE) of 5.10 years, demonstrating marked performance improvements, particularly within the challenging adolescent cohort. The model maintained strong generalization on the unseen dataset (MAE = 4.19 years) and exhibited superior longitudinal consistency (Mean Absolute Difference Error, MAdE = 3.68) and excellent test–retest reliability (Intraclass Correlation Coefficient, ICC = 0.994). By integrating a large-scale, heterogeneous lifespan dataset with a hybrid architecture that effectively captures both local structural details and global long-range interactions, our study provides a precise, generalizable, and reliable framework for brain age estimation.