Learning neuroimaging models from health system-scale data

2026-02-06 · Source: Machine learning : nature.com subject feeds · Field: Health & Wellbeing — Medical Devices & Health Technology, Clinical Care & Medical Practice, Medical Specialties & Subspecialties · Depth: Expert, extended

Summary

Prima, an AI foundation model for neuroimaging, was developed using over 220,000 MRI studies from a large academic health system to address rising demand and physician burnout in neurological disease evaluation. This model utilizes a hierarchical vision architecture to generate general and transferable MRI features, supporting real-world clinical MRI inputs. In a one-year health system-wide study involving 29,431 MRI scans, Prima achieved a mean diagnostic Area Under the Curve (AUC) of 92.0% across 52 radiologic diagnoses, outperforming other general and medical AI models. The system also provides explainable differential diagnoses, worklist prioritization for radiologists, and clinical referral recommendations, while demonstrating algorithmic fairness across diverse patient groups. The model parameters and code, implemented in Python with PyTorch and Transformers, are publicly available under an MIT license for investigational use.

Key takeaway

For AI scientists and neuroradiology teams developing diagnostic tools, Prima demonstrates that training on health system-scale MRI data significantly boosts diagnostic accuracy and clinical utility. Your efforts should focus on integrating similar large-scale, real-world datasets and hierarchical vision architectures to create robust, explainable, and fair AI models that can genuinely alleviate clinical burdens and improve patient outcomes.

Key insights

Health system-scale data enables AI foundation models to achieve superior, fair, and explainable neuroimaging diagnostics.

Principles

• Large-scale, real-world data improves AI diagnostic accuracy.
• Hierarchical vision architectures enhance MRI feature transferability.
• Algorithmic fairness is achievable across sensitive patient groups.

Method

Prima uses a hierarchical vision transformer trained with a CLIP objective on 3D MRI volumes and summarized radiology reports, followed by transfer learning with an MLP for diagnostic prediction.

In practice

• Utilize GPT-3.5-turbo for radiology report summarization.
• Implement VQ-VAE for MRI volume tokenization.
• Apply LIME for explainable AI predictions in clinical context.

Topics

• Neuroimaging AI
• AI Foundation Models
• Medical Diagnostics
• Algorithmic Fairness
• Health System Data

Code references

• MLNeurosurg/Prima

Best for: Computer Vision Engineer, AI Scientist, Research Scientist, AI Engineer, Machine Learning Engineer, AI Researcher

Related on AIssential

• AI reads brain MRIs in seconds and flags emergencies — A new AI system rapidly interprets brain MRIs, achieving high accuracy and prioritizing urgent neurological conditions.
• Scaling Generative Foundation Models for Chest Radiography with Rectified Flow Transformers — A 1.3B-parameter generative foundation model for chest radiographs achieves expert-indistinguishable synthesis and controllable editing.
• Improving clinical interpretability of linear neuroimaging models through feature whitening — A novel whitening method improves neuroimaging model interpretability by decorrelating anatomically informed brain regions.
• AI-powered evaluation of dementia severity based on clinical data and visual scoring systems (MTA, ERICA, GCA) from MRI — AI-powered visual scoring and clinical data integration improves dementia severity evaluation, reducing expert dependency.
• If you have enough data and enough compute, general learning machines that “discover” patterns tend to outperform systems where humans try to hard-code expertise. — General learning systems with vast data and compute outperform human-designed expert AI, a "bitter lesson" now impacting surgical AI.
• Improving breast cancer screening workflows with machine learning — Google Research, in partnership with NHS organizations, conducted the Artificial Intelligence in Mammography Screening (AIMS) study, publishing two companion studies in Nature Cancer to evaluate an AI…

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine learning : nature.com subject feeds.