AI Gives a Clearer Picture of Functional MRI Brain Data Using Advanced Denoising Techniques

Functional magnetic resonance imaging, better known as fMRI, has been one of the most important tools in modern neuroscience for over three decades. It allows researchers to observe brain activity noninvasively by tracking changes in blood flow related to neural activity. Despite its widespread use and tens of thousands of studies published every year, fMRI has always faced a persistent and frustrating limitation: noise.

Now, researchers at Boston College report a major breakthrough. In a study published in Nature Methods, the team demonstrates that artificial intelligence, specifically generative AI, can dramatically improve the quality of fMRI data by removing noise far more effectively than existing methods. Their new approach, called DeepCor, offers a clearer and more reliable picture of how the human brain functions and how it may malfunction in disease.

Why Noise Has Always Been a Problem in fMRI

fMRI works by detecting changes in oxygenated blood, which indirectly reflect neural activity. While powerful, the technique is extremely sensitive to unwanted signals. Even small head movements, natural heartbeat rhythms, breathing, or tiny fluctuations in the MRI machine itself can distort measurements.

These distortions mix with genuine brain responses, making it difficult to tell what is meaningful neural activity and what is simply noise. Over time, researchers have developed several denoising techniques to address this issue, but most rely on simplifying assumptions about how noise behaves. As a result, important brain signals can still be lost or blurred.

This limitation has long restricted how confidently scientists can interpret fMRI results, especially when studying subtle brain responses or clinical conditions.

How AI Changes the Equation

The Boston College research team took a different approach by turning to generative artificial intelligence. Their method, DeepCor, uses contrastive autoencoders, a type of deep learning model designed to learn patterns in complex data.

Instead of trying to model noise directly, DeepCor compares two kinds of brain regions:

• Regions that contain neurons, where meaningful brain signals are expected
• Regions without neurons, such as ventricles, which should contain no neural activity but still capture noise

Because noise affects both types of regions in similar ways, the AI learns to identify patterns shared between them. By removing these shared patterns, DeepCor isolates what truly matters: the unique neural signals from regions that contain neurons.

This contrastive strategy allows the system to separate signal from noise more accurately than traditional methods, without relying on overly simplistic assumptions.

A Bigger Improvement Than Expected

When the researchers tested DeepCor, the results exceeded even their most optimistic expectations.

On simulated datasets, designed to closely resemble real fMRI data, DeepCor outperformed existing state-of-the-art methods by more than 200 percent. In one particularly striking comparison, it improved clarity by 339 percent over conventional approaches.

The method was also tested on real human fMRI data, where it was compared to a widely used technique known as CompCor. In these real-world conditions, DeepCor removed noise from face-related brain responses 215 percent more effectively than CompCor.

Such improvements are not incremental tweaks. They represent a major leap in how clean and interpretable fMRI data can be.

Why This Matters for Brain Research

Cleaner fMRI data has far-reaching implications. With better noise removal, researchers can:

• Detect subtle brain responses that were previously hidden
• Improve studies of perception, cognition, and behavior
• Gain more reliable insights into neurological and psychiatric disorders

Because fMRI is used so widely across neuroscience and clinical research, improvements at the data-processing level can influence thousands of studies and potentially reshape conclusions drawn from existing datasets.

The researchers were particularly surprised by how large the gains were. They initially expected improvements in the range of 10 to 50 percent. Achieving gains above 200 percent highlighted just how powerful generative AI can be when applied thoughtfully to scientific data.

The DeepCor Team and Their Work

The study was led by Stefano Anzellotti, Associate Professor of Psychology at Boston College, along with Aidas Aglinskas, a postdoctoral researcher, and Yu Zhu, who was an undergraduate student at the time of the research.

Their work involved extensive testing on both synthetic and real fMRI datasets, ensuring that DeepCor was not just theoretically promising but also practically useful.

Importantly, DeepCor does not require massive training datasets across many subjects. It can be applied flexibly, making it suitable for a wide range of experimental designs, including task-based and resting-state fMRI studies.

Making the Method Widely Accessible

Looking ahead, the research team is focused on two major goals. First, they aim to make DeepCor easy to access and use so that researchers around the world can integrate it into their workflows without specialized expertise in AI.

Second, they plan to apply DeepCor to large public fMRI datasets. By denoising these datasets at scale, the broader neuroscience community could benefit from cleaner data almost immediately, potentially leading to faster progress and more reliable discoveries.

A Closer Look at fMRI and AI in Neuroscience

This research also reflects a broader trend in neuroscience: the growing role of machine learning and AI in data analysis. As brain imaging technologies generate increasingly complex datasets, traditional statistical tools often struggle to keep up.

AI models excel at identifying patterns in high-dimensional data, making them especially well suited for imaging tasks. However, concerns remain about interpretability and reliability. DeepCor addresses these concerns by grounding its approach in biological reasoning, explicitly leveraging known differences between neural and non-neural brain regions.

This balance between advanced computation and neuroscience insight is likely to define the next generation of brain imaging research.

What This Could Mean for Clinical Applications

While the current study focuses on research use, improved fMRI data quality could eventually influence clinical practice. Better signal clarity may help researchers identify biomarkers for conditions such as depression, schizophrenia, neurodegenerative diseases, or developmental disorders.

Although clinical translation will require further validation, DeepCor represents an important step toward making fMRI a more precise and reliable tool for understanding the human brain.

A Clearer Brain Picture, Powered by AI

DeepCor shows that artificial intelligence is not just accelerating analysis, but fundamentally improving data quality. By tackling one of fMRI’s oldest challenges, this work opens the door to more accurate brain maps, deeper insights into neural function, and a stronger foundation for future discoveries.

As AI continues to merge with neuroscience, methods like DeepCor suggest that some of the brain’s long-standing mysteries may soon come into sharper focus.

Research paper:
https://doi.org/10.1038/s41592-025-02967-x