Brain Imaging Dataset Advances Hearing Health Research

A new, publicly available brain imaging dataset contains over 260,000 MRI scans from more than 55,000 individuals. Published in *Scientific Data* in 2026, this resource aims to accelerate the development of artificial intelligence tools for analyzing brain structure. The dataset’s scale and diversity could have significant downstream implications for understanding neurological conditions related to hearing and sound perception, such as tinnitus, misophonia, and hyperacusis.

### A Massive, Heterogeneous Brain Imaging Archive

Led by Stefano Cerri and Asbjørn Munk from the University of Copenhagen and the Pioneer Centre for AI, the international team assembled the dataset, named FOMO260K. It represents one of the largest collections of its kind intended for machine learning research. The 260,927 scans come from 77,589 separate MRI sessions. Critically, the dataset is heterogeneous, meaning it combines images from many different scanners, protocols, and patient populations. It includes both high-resolution research scans and routine clinical images, the latter often containing large brain anomalies like tumors or strokes. The researchers applied only minimal preprocessing to preserve the raw characteristics of the original data, which better prepares AI models for the messy reality of clinical practice.

### Methodology: Built for Self-Supervised Learning

The core purpose of FOMO260K is to serve as a training ground for self-supervised learning (SSL) models. Unlike traditional supervised AI that needs millions of manually labeled images (e.g., “this scan shows a tumor”), SSL algorithms learn by finding patterns and relationships within the data itself. A common SSL task is to train a model to reconstruct a missing part of an image or to determine if two differently processed views came from the same original scan. By learning a rich, generalized understanding of normal and abnormal brain anatomy from a quarter-million examples, these pre-trained models can then be efficiently adapted, or “fine-tuned,” for specific tasks with much smaller, labeled datasets. The authors provide companion code and pre-trained models to help other researchers start this work.

### Potential Implications for Hearing and Sound Disorder Research

For researchers studying tinnitus, hyperacusis, and misophonia, large-scale neuroimaging resources are vital. These conditions are believed to involve complex changes in brain networks responsible for sound processing, attention, and emotion. Pinpointing reliable structural brain markers has been challenging, partly due to the limited size and variety of most single-site imaging studies. FOMO260K changes this dynamic.

A model pre-trained on such a vast and varied collection learns a more robust representation of the human brain. It becomes better at detecting subtle, condition-specific deviations that might be missed by models trained on smaller, cleaner datasets. For instance, fine-tuning such a model on a cohort of patients with tinnitus could potentially identify consistent, yet previously overlooked, structural changes in the auditory cortex or limbic system. This aligns with research exploring the shared neurobiology between PTSD and tinnitus, where overlapping brain networks are implicated.

Furthermore, the inclusion of clinical scans with pathologies is a major strength. It can help researchers distinguish brain changes unique to a condition like misophonia from those caused by unrelated, co-occurring neurological issues. This improves diagnostic specificity. Understanding these neural substrates is a step toward more objective measures for disorders currently defined by subjective report, similar to efforts exploring the link between migraine and auditory dysfunction.

### Lowering Barriers and Accelerating Discovery

The public release of FOMO260K is a significant move toward democratizing medical AI research. By providing open access to the dataset and code, the team lowers the entry cost for academic labs and clinicians who wish to develop diagnostic or prognostic tools. This could accelerate the pace of discovery across neuroscience, including hearing health.

Instead of spending years and considerable resources collecting their own massive imaging dataset, a lab focused on hyperacusis can start with a sophisticated pre-trained model from FOMO260K. They can then fine-tune it using their specialized, smaller dataset of patient scans. This approach makes advanced computational analysis more accessible, potentially leading to faster identification of patient subtypes, better prediction of treatment outcomes, and a deeper understanding of the continuum between conditions like misophonia, stress, and hearing health.

The dataset also establishes a common benchmark. Researchers can now compare the performance of different AI models on the same massive scale, ensuring progress is measurable and based on a standardized foundation.

**Source Paper:** Cerri S, Munk A, Llambias SN, et al. A large-scale heterogeneous 3D magnetic resonance brain imaging dataset for self-supervised learning. *Sci Data*. 2026. DOI: 10.1038/s41597-026-07688-0. PMID: 42420299.