⚡ electrical engineering

A Two-Stage Multi-Modal MRI Framework for Lifespan Brain Age Prediction

This paper presents a novel two-stage, multi-modal MRI framework that overcomes the limitations of existing narrow-range, single-modality approaches by first classifying subjects into six developmental stages and then estimating their age within those stages to enable accurate, lifespan-spanning brain age prediction.

Original authors: Dingyi Zhang, Ruiying Liu, Yun Wang

Published 2026-04-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Dingyi Zhang, Ruiying Liu, Yun Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your brain is like a house. As you grow from a baby into an elderly person, the "house" changes its architecture. The walls get thicker, the wiring (nerves) gets more organized, and eventually, some parts might start to wear down.

Doctors and scientists want to know how "old" this house really is based on its current condition. This is called Brain Age Prediction. If your brain looks like it belongs to a 50-year-old, but you are only 30, that's a warning sign that your brain might be aging faster than it should.

However, the current tools for checking this "brain age" have some big problems:

They usually only work for specific age groups (like just adults or just babies), not the whole life.
They often look at only one type of picture (like a black-and-white photo) instead of a full color 3D scan.
If a picture is missing a piece, the tool often breaks.

This paper introduces a new, smarter system called a "Two-Stage Multi-Modal Framework" that fixes all these issues. Here is how it works, explained with simple analogies:

1. The "All-Seeing" Camera (Multi-Modal)

Most old systems take just one photo of the brain. This new system takes three different types of photos at the same time:

T1w & T2w: Like taking a high-resolution photo of the brain's structure (the walls and rooms).
FA (Fractional Anisotropy): Like taking a photo of the "wiring" inside the walls to see how well the signals travel.

The Magic Trick: If the hospital only has one of these photos (maybe the wiring photo is missing), the system doesn't crash. It uses a "Late Fusion" strategy. Think of it like a panel of judges. If Judge A (T1w) is present but Judge B (FA) is absent, the panel still makes a decision based on who is there. It doesn't need all judges to be present to give a verdict.

2. The Two-Stage Process (The "Sort and Measure" System)

Instead of trying to guess a person's exact age immediately (which is hard when comparing a fetus to an elderly person), the system does it in two steps:

Stage 1: The "School Grade" Classifier
Imagine a teacher trying to guess a student's age. Instead of guessing "14 years, 3 months," the teacher first asks: "Is this student in Kindergarten, Elementary, High School, or College?"

The system first sorts the brain into one of six life stages: Fetal, Newborn, Infant, Child, Adult, or Elderly.
It uses a special "Mixture of Experts" (MoE) network. Think of this as a team of specialists. One expert only knows about babies, another only knows about adults. The system automatically picks the right expert to look at the brain and say, "Ah, this is definitely a Child."

Stage 2: The "Fine-Tuned" Estimator
Once the system knows the brain is a "Child," it doesn't try to guess the age of an adult. It switches to a specialist who only looks at children.

This specialist then gives a very precise age estimate (e.g., "This child is 4.5 years old").
Because the system isn't confused by the huge differences between a fetus and an elderly person, it is much more accurate.

3. Why This is a Big Deal

The researchers tested this system on nine different datasets covering the entire human lifespan, from unborn babies to the elderly.

The Result: It was the most accurate system ever tested.
The Analogy: If other systems were like a generic ruler that tried to measure a grain of sand and a skyscraper with the same markings, this new system is like having a microscope for the sand and a laser measure for the skyscraper, and knowing exactly which tool to use based on what you are looking at.

Summary

This paper presents a smart, flexible AI that can look at a brain at any age in life. It uses multiple types of brain scans, handles missing data gracefully, and uses a "sort-then-measure" strategy to give the most accurate "biological age" possible. This helps doctors spot health issues earlier, whether the patient is a newborn or a senior citizen.

1. Problem Statement

Brain age prediction, which estimates biological age from MRI scans to assess neuroanatomical health, faces three critical limitations in current state-of-the-art methods:

Restricted Lifespan Coverage: Existing models typically focus on narrow age ranges (e.g., only adults or only children) rather than the full human lifespan from fetal development to old age.
Single-Modality Reliance: Most approaches use only T1-weighted (T1w) images, failing to capture complementary macro-structural (cortical) and micro-structural (white matter) information provided by other modalities like T2-weighted (T2w) and Fractional Anisotropy (FA).
Robustness to Missing Data: Real-world clinical settings often suffer from incomplete multimodal data. Existing models usually require all modalities for training and inference, lacking robustness when specific modalities are missing.

2. Methodology

The authors propose a Two-Stage Multi-Modal Framework designed to handle the full lifespan and missing modalities through a late-fusion strategy.

A. Data Preprocessing & Age Unification

Modalities: T1w, T2w, and FA maps.
Preprocessing: N4 bias correction, skull stripping (using Synthstrip), and 1mm isotropic resampling.
Age Mapping: To unify the age axis, fetal gestational weeks ( $w$ ) are mapped to a continuous scale where $y = w - 40$ . This creates a continuous timeline from prenatal stages to elderly.

B. Stage 1: Age Stage Classification

The human lifespan is partitioned into six developmental stages:

Fetal (20–40 weeks)
Neonatal (0–12 weeks postnatal)
Infant (3–24 months)
Child (2–17 years)
Adult (18–65 years)
Elderly (>65 years)

Architecture: A shared Masked Autoencoder (MAE) backbone encodes each modality independently.
Modality MoE: A Mixture of Experts (MoE) with hard routing processes the latent representations. Each expert specializes in a specific MRI modality.
Fusion: Predictions from available modalities are aggregated to determine the most probable stage ( $\hat{s}$ ).
$\hat{s} = \arg \max_s \sum_m p(s | x_m)$

C. Stage 2: Brain Age Regression

Once the developmental stage is predicted, a regression network estimates the precise brain age ( $\hat{y}$ ) conditioned on $\hat{s}$ .

Age Stage MoE: A second MoE layer is introduced where each expert specializes in a specific lifespan stage to capture stage-specific macro/microstructural patterns.
Transfer Learning: The MAE backbone and Modality MoE from Stage 1 are reused and fine-tuned.
Late Fusion: Modalities are processed independently, and their regression outputs are aggregated at inference. This allows the model to function even if some modalities are missing, without retraining.

3. Key Contributions

First Lifespan Framework: The first model to span the complete human lifespan (fetal to elderly) in a unified architecture.
Robust Multi-Modal Integration: Utilizes a late fusion strategy integrating T1w, T2w, and FA. It does not require multi-modal training data and naturally handles missing modalities during inference.
Two-Stage MoE Architecture: Combines coarse-grained stage classification with fine-grained regression, leveraging Mixture of Experts to handle diverse, age-specific, and modality-specific patterns effectively.

4. Experimental Results

The model was evaluated on nine datasets (5 in-domain for training/validation, 4 out-of-domain for testing), covering the full age spectrum.

Performance Metrics: Mean Absolute Error (MAE) and Standard Deviation (STD) in years.
In-Domain Results: The multi-modal variant (Ours-M) achieved the best overall performance, outperforming baselines like SFCN, ORDER, and TSAN.
- Example: In the Fetal stage, Ours-M achieved an MAE of 0.01 years (vs. 0.14–0.87 for baselines).
- Example: In the Elderly stage, Ours-M achieved an MAE of 5.77 years (vs. 6.30–8.92 for baselines).
Out-of-Domain Generalization: The model demonstrated superior generalization to unseen datasets (e.g., ADNI, ABCD).
- Ours-M reduced MAE by 69% in out-of-domain settings compared to existing baselines.
- Notably, in early developmental stages (fetal/neonatal), prediction errors were close to zero, indicating high sensitivity to rapid developmental changes.
Ablation: The multi-modal version (Ours-M) consistently outperformed the single-modal version (Ours-S), confirming the value of integrating complementary structural information.
Stage Classification: The first-stage classifier achieved 88.18% accuracy on out-of-domain data, significantly reducing extreme misclassifications (e.g., misclassifying neonates as adults) compared to single-modal approaches.

5. Significance

This work represents a significant advancement in neuroimaging AI by:

Bridging Developmental Gaps: Providing a unified tool to study brain maturation and aging across the entire human lifespan, which is crucial for understanding developmental disorders and neurodegenerative diseases.
Clinical Applicability: The ability to handle missing modalities makes the framework highly practical for real-world clinical environments where complete MRI protocols are not always feasible.
Biological Insight: The use of MoE and multi-modal data allows the model to learn distinct biological signatures for different life stages, offering a more nuanced view of brain health than single-stage, single-modality models.

The paper concludes that this two-stage, multi-modal approach sets a new standard for robust, lifespan-spanning brain age estimation, reducing prediction errors by up to 69% in challenging generalization scenarios.