Artificial intelligence applied to medical imaging has experienced remarkable advances in recent years, but general-purpose foundation models have stumbled upon a recurring obstacle: the lack of quality clinical neuroimaging data. While models like GPT are trained on vast Internet corpora, magnetic resonance imaging (MRI) and computed tomography (CT) rarely appear on the public web because they contain identifiable facial features. This data gap has limited the performance of general-purpose models in brain diagnosis tasks. Facing this challenge, a team at the University of Michigan has developed NeuroVFM, a visual foundation model for neuroimaging that learns directly from uncured clinical practice, without the need for human labels or paired radiology reports. Its core architecture, Vol-JEPA, extends JEPA approaches to three-dimensional volume, predicting latent representations instead of pixels. The results, published in Nature Medicine, show a significant improvement over previous models, with an AUROC above 92% for both CT and MRI. This achievement not only represents a scientific breakthrough but also opens the door to new applications in triage, report generation, and AI-assisted diagnosis, all at a computational cost far lower than proprietary models like GPT-5.
But beyond the laboratory, implementing solutions like NeuroVFM in real healthcare environments requires a deep reflection on the necessary technological infrastructure. Having a powerful model is not enough: custom software is needed to integrate these models into clinical workflows, ensuring security, scalability, and regulatory compliance. Companies like Q2BSTUDIO, specialized in software development, artificial intelligence, and cybersecurity, are perfectly positioned to help hospitals and research centers deploy such models efficiently and safely. The combination of cutting-edge foundation models with a robust technological platform is the key to transforming the promise of AI into real impact on patient health.
In this article we will explore in detail how NeuroVFM works, what sets it apart from other approaches, what results it has achieved in clinical practice, and how technology companies can facilitate its adoption through cloud computing services, AI agents, and business intelligence solutions.
The problem of lack of clinical data in generalist models is well known. Current foundation models, such as GPT-4o, Claude, or Gemini, are predominantly trained on text and images extracted from the Internet. However, clinical neuroimaging (MRI and CT) images are protected by privacy regulations such as HIPAA in the United States or GDPR in Europe, as they contain identifiable biometric information. As a result, these models have rarely seen a real brain scan during training, drastically limiting their ability to interpret neurological pathologies.
The University of Michigan team addressed this limitation through what they call 'health system learning.' Instead of relying on curated, manually labeled datasets, they used 5.24 million clinical MRI and CT volumes obtained over two decades of routine care at Michigan Medicine. These data were not cleaned or annotated; they were simply extracted from the usual clinical workflow. The model learned to represent brain anatomy and anomalies without requiring radiologist reports or diagnostic labels. This approach avoids the bottleneck of having to pair each image with a textual report and eliminates the disease-selection bias that affects traditional narrow classifiers.
The model architecture is based on Vol-JEPA, a three-dimensional extension of the JEPA (Joint Embedding Predictive Architecture) methods previously developed by the Meta AI team. While I-JEPA operates on 2D images and V-JEPA on videos, Vol-JEPA processes complete 3D volumes. The process begins by tokenizing each volume into non-overlapping patches of 4x16x16 voxels. The volume is then split into a small visible portion (context) and a larger masked portion (target). A student encoder processes the context patches, and a predictor combines the context latent representations with the target position encodings to predict the latent representations of the masked region. A teacher encoder, which is an exponential moving average of the student, generates the actual target representations. The loss function is a smooth L1 between the predicted and teacher representations, with gradients stopped through the teacher.
Masking is selective: it focuses on the head region using precomputed masks. For MRI, 25% context is used; for CT, 20%, with 20% patch dropout. This forces the encoder to model shared neuroanatomy rather than learning shortcuts based on image background. The result is a model with 85.8 million parameters in its base version and 21.7 million in the small version, trained in less than 1,000 GPU hours, seven times faster than a 3DINO baseline, and capable of processing batches 16 times larger with the same memory.
Quantitative results are impressive. The model achieved a macro-averaged AUROC of 92.68% on CT and 92.49% on MRI over 156 diagnostic tasks (74 MRI and 82 CT). It outperformed all baselines, including HLIP (trained with report supervision), NeuroMAE (voxel reconstruction), PRIMA, DINOv3, and BiomedCLIP. The difference with HLIP, which shares the same training data but uses textual supervision, was -0.98 percentage points, demonstrating that latent prediction is superior to report supervision. Compared to NeuroMAE, the difference was -1.55 points, showing that predicting latent representations is more effective than reconstructing pixels.
Beyond diagnosis, NeuroVFM enables advanced report generation and triage tasks. The researchers combined the frozen NeuroVFM encoder with a large language model, Qwen3-14B, in a LLaVA-1.5 style architecture, creating NeuroVFM-LLaVA. This system generates structured findings from neuroimaging volumes and, through a reasoning model, assigns acuity levels: unremarkable, routine, or urgent. The results outperformed GPT-5 in triage accuracy (92.6% balanced accuracy vs. 71.2%) and in critical-finding miss rate (13.5% vs. 50.3%). Additionally, inference cost was more than 24 times lower and carbon footprint more than 23 times lower.
However, the model is not perfect. In a silent one-week prospective study with 1,155 patients, NeuroVFM achieved a sensitivity of 86.5%, meaning it missed 21 of 155 critical findings. The authors emphasize that it should be used as clinical decision support, not as autonomous screening. Furthermore, the model is not FDA-approved and its weights are distributed under a CC-BY-NC-SA-4.0 license, limiting commercial use. There is also a risk of bias related to architecture, dataset, and learning objective, as results come from a single academic health system.
From a business perspective, implementing models like NeuroVFM in a hospital or diagnostic center requires adequate technological infrastructure. This is where companies like Q2BSTUDIO can make a difference. We offer cloud services on Azure and AWS that allow deploying AI models at scale with high availability and security. Cybersecurity is especially critical in healthcare, where patient data is protected by strict regulations; therefore, we provide pentesting and security auditing services to ensure compliant deployment. Additionally, our AI agents can automate triage workflows, allowing the NeuroVFM model to integrate seamlessly into hospital information systems.
Another essential aspect is the visualization and analysis of results. Models generate a large amount of data that must be interpreted by radiologists. Through Business Intelligence and Power BI services, we can create interactive dashboards showing model predictions, acuity levels, and temporal trends, facilitating clinical decision-making and resource management. Likewise, custom software development allows adapting the user interface to each hospital's specific needs, integrating AI naturally into the radiologist's workflow.
In summary, NeuroVFM represents a milestone in clinical neuroimaging by demonstrating that a model trained exclusively on uncured real-world data can outperform models trained with textual supervision or pixel reconstruction. Its efficient and scalable Vol-JEPA architecture opens the door to a new generation of AI-assisted diagnostic tools. However, the path to widespread clinical adoption requires more than a good model: it needs a solid technological platform that ensures security, scalability, and integration. At Q2BSTUDIO we offer exactly that: custom software development, cloud computing, cybersecurity, AI agents, and BI to transform innovation into real impact. AI in healthcare is not the future; it is the present, and we are ready to accompany organizations on this journey.
