Understanding Sources of Demographic Predictability in Brain MRI via Disentangling Anatomy and Contrast

This paper proposes a disentangled representation learning framework for brain MRI to demonstrate that demographic predictability primarily stems from anatomical variation rather than acquisition-dependent contrast, highlighting the need for targeted mitigation strategies that address these distinct sources to ensure robust bias reduction.

The Gist

The Big Question: Why Can AI Guess Your Age and Race from a Brain Scan?

Imagine you walk into a doctor's office, and they take a picture of your brain. A few years ago, you might have thought, "This picture just shows my brain." But today, advanced AI can look at that same picture and guess your age, your sex, and even your race with surprising accuracy.

This is a problem. If an AI can guess your race just by looking at a brain scan, it might accidentally learn to treat patients differently based on their race, even if the doctor didn't ask for that information. This is called bias.

But here is the mystery: Where does this "clue" come from?

• Is it because your brain physically looks different (anatomy)?
• Is it because the machine that took the picture was set up differently (contrast)?
• Or is it a mix of both?

Until now, nobody knew for sure because the two things were "entangled" (mixed together) like a bowl of spaghetti. You couldn't separate the noodles from the sauce.

The Solution: The "Magic Filter"

The researchers at King's College London built a special "magic filter" (a computer program) to untangle the spaghetti. They used a technique called Disentangled Representation Learning.

Think of a brain MRI scan like a photograph of a house:

1. The Structure (Anatomy): The shape of the walls, the size of the rooms, the layout of the kitchen. This is the "real" house.
2. The Lighting (Contrast): Is the photo taken in bright sunlight? Is it a black-and-white photo? Is it taken with a vintage camera or a modern one? This is the "acquisition" part.

In a normal photo, you can't tell the difference between the house's shape and the lighting. The researchers created a tool that splits the photo into two separate files:

• File A (Anatomy): Shows the house shape, but removes all the lighting effects. It looks the same whether the sun is shining or it's raining.
• File B (Contrast): Shows only the lighting and camera style, but blurs out the house shape so you can't see the rooms anymore.

What They Discovered

The researchers trained three different AI "detectives" to guess age, sex, and race:

1. Detective Raw: Looked at the original, unfiltered photo.
2. Detective Structure: Looked only at File A (the house shape).
3. Detective Lighting: Looked only at File B (the lighting).

Here is what they found:

1. The Shape is the Main Clue (The House Matters Most)

Detective Structure was almost as good as Detective Raw.

• The Analogy: Even if you take away the lighting and just look at the blueprint of the house, the AI can still guess your age and sex.
• The Science: This means that biological differences in brain size and shape are the primary reason AI can guess your demographics. For example, older brains naturally shrink, and male and female brains have slightly different shapes. This is real biology, not a glitch.

2. The Lighting Still Hints at Something (The Camera Matters Too)

Detective Lighting wasn't perfect, but it wasn't useless either. It could still guess your demographics better than random chance.

• The Analogy: If you only look at the "lighting" file, the AI can tell you, "Hey, this photo looks like it was taken with a specific type of camera in a specific city."
• The Catch: This clue is fragile. If you train the AI on photos taken in London and then test it on photos taken in New York, the "Lighting Detective" fails completely. The clues it learned were specific to that one hospital's machine settings, not a universal truth.

Why This Matters for the Future

This study is a wake-up call for how we fix AI bias.

• The Old Way: People thought, "If we just fix the lighting (standardize the machine settings), the bias will go away."
• The New Reality: The paper says, "No, that won't work." Even if you fix the lighting perfectly, the AI will still guess your race and age because of the shape of the brain.

The Takeaway:
To make AI fair, we can't just fix the cameras. We have to understand that the AI is learning from real biological differences.

• If the AI is guessing your age because your brain is shrinking, that's actually useful medical information!
• But if the AI is guessing your race because of subtle shape differences that correlate with race, we need to be very careful not to let that turn into discrimination.

Summary in One Sentence

The AI isn't just "cheating" by looking at machine settings; it's actually reading the real physical differences in our brains, so fixing bias requires a much smarter approach than just standardizing the equipment.

Technical Summary

1. Problem Statement

Demographic attributes (age, sex, race) can be predicted from medical images, raising concerns about bias in clinical AI systems. In brain MRI, the source of this "demographic signal" is ambiguous. It may arise from:

• Anatomical Variation: True biological differences (e.g., brain volume, morphology).
• Acquisition-Dependent Contrast: Technical artifacts caused by scanner hardware, site-specific protocols, and sequence parameters (TR, TE).
• Interaction: A combination of both.

Current mitigation strategies often fail because they treat the demographic signal as a monolithic issue. Without disentangling these sources, interventions risk either failing to address the root cause or inadvertently removing clinically relevant information. The core challenge is that in raw MRI data, anatomy and contrast are inherently entangled.

2. Methodology

The authors propose a controlled framework using disentangled representation learning to decompose brain MRI into two distinct components:

1. Anatomy-focused representation ($z_{anat}$): Captures structural/geometric information while suppressing acquisition-dependent intensity variations.
2. Contrast-only embedding ($z_{contrast}$): Captures acquisition-dependent characteristics (scanner/protocol effects) while minimizing anatomical content.

Framework & Models

The study leverages two pre-trained models:

• MR-CLIP: A metadata-guided contrastive learning framework that aligns MRI images with DICOM acquisition parameters (without manual labels) to learn contrast embeddings.
• DIST-CLIP: Builds on MR-CLIP by introducing an explicit anatomical pathway. It uses an anatomy mapper to project images into $z_{anat}$, enforcing acquisition invariance via patch-level contrastive loss. This ensures corresponding anatomical patches remain similar regardless of acquisition differences.

Experimental Design

The authors trained predictive models for Age (regression), Sex, and Race (classification) on three distinct input types:

1. Raw Images ($x$): The full, entangled input.
2. Anatomy Representations ($z_{anat}$): Structural only.
3. Contrast Embeddings ($z_{contrast}$): Acquisition characteristics only.

Datasets & Evaluation

• Datasets: OASIS-3, ADNI, and HCP (spanning different ages, scanner vendors, and protocols).
• Preprocessing: Rigid registration to MNI152 atlas, skull stripping (SynthStrip).
• Metrics: Mean Absolute Error (MAE) for age; Balanced Accuracy (BalAcc) for sex and race (to handle class imbalance).
• Evaluation Strategy:
  • Within-dataset: To quantify the relative contribution of structure vs. contrast.
  • Cross-dataset: To test generalization (e.g., train on OASIS, test on HCP) and determine if the signal is stable biological variation or site-specific artifact.
  • Multi-sequence: Analysis across T1w, T2w, and FLAIR sequences.

3. Key Contributions

1. Controlled Framework: A novel methodology to explicitly separate anatomical structure from acquisition-dependent contrast in brain MRI using disentangled representation learning.
2. Source Attribution: Empirical evidence across three datasets and multiple sequences demonstrating that anatomical structure is the dominant carrier of demographic information.
3. Generalization Insights: Discovery that while contrast embeddings retain a systematic demographic signal, it is dataset-specific and fails to generalize across sites, whereas anatomical signals are more robust.

4. Key Results

A. Within-Dataset Performance

• Raw Images: Achieved the highest performance (e.g., Sex BalAcc 0.92–0.96).
• Anatomy ($z_{anat}$): Largely preserved the performance of raw images. For example, Sex prediction on OASIS was 0.92 (Raw) vs. 0.93 (Anat). This indicates that the majority of the demographic signal is embedded in structural variation.
• Contrast ($z_{contrast}$): Retained non-trivial predictive power despite minimizing anatomy.
  • Sex: 0.71 (OASIS), 0.81 (HCP).
  • Race: 0.71 (OASIS), 0.70 (HCP).
  • Age: MAE of 6.28 years (OASIS) and 2.77 years (HCP).
  • Conclusion: Acquisition characteristics alone carry a measurable, systematic demographic signal.

B. Cross-Dataset Generalization

• Anatomy: Demonstrated robust generalization. When moving from OASIS to HCP, Sex prediction improved from 0.65 (Raw) to 0.82 (Anat).
• Contrast: Frequently collapsed to chance levels (0.50) when tested on a different dataset.
• Interpretation: The demographic signal in contrast embeddings is heavily tied to specific site/protocol distributions and does not transfer well. Anatomical signals are more biologically stable across domains.

C. Multi-Sequence Analysis

• The dominance of anatomical structure held true across T1w, T2w, and FLAIR sequences.
• Contrast embeddings remained informative but weaker across all sequences, confirming the findings are not specific to a single modality.

5. Significance and Implications

• Bias Mitigation Strategy: Effective bias reduction cannot rely solely on acquisition harmonization (e.g., intensity normalization) because a significant portion of the demographic signal is biological (anatomical). Conversely, indiscriminately suppressing anatomical variation could remove clinically meaningful data.
• Targeted Intervention: Mitigation strategies must be source-aware.
  • For acquisition-driven bias: Harmonization and domain adaptation are effective.
  • For anatomical-driven bias: Strategies must address the underlying biological correlations (e.g., via representation debiasing or causal modeling) without compromising diagnostic utility.
• Clinical Safety: Understanding the origin of the signal is a prerequisite for ensuring that bias reduction does not inadvertently degrade model performance on specific clinical tasks (e.g., Alzheimer's detection).

6. Limitations

• Data Imbalance: The racial composition of the datasets is imbalanced (e.g., ADNI is 94.8% White), which limits the statistical reliability of race prediction results despite the use of balanced accuracy.
• Scope: The study focuses on brain MRI; findings may not directly translate to other modalities like chest X-rays or cardiac MRI without further investigation.

Conclusion

The paper establishes that demographic predictability in brain MRI is primarily rooted in anatomical variation, with a secondary, site-specific contribution from acquisition-dependent contrast. This distinction is critical for developing fair, robust, and clinically safe AI systems, as it dictates that bias mitigation must address both biological and technical sources of variation.