New AI Trained on 2 Million Medical Scans Could Speed Diagnosis

Doctors who read scans work in three dimensions — slicing through a body image like a loaf of bread to find a tumor or a blocked artery. Most medical AI today is built for one narrow job, like spotting a single disease, and it often stumbles when it meets a new hospital's machines or patients. A large research team has now released nnFoundation, a "foundation model" — a general-purpose AI that learns broadly first, then gets adapted to many jobs. It studied 2.1 million CT, MRI and PET scans from 125 different datasets.

The team tested it on 108 tasks: outlining organs, detecting problems, sorting scans into categories, writing draft reports, and finding similar past cases. nnFoundation beat both older 3D medical AI models and AI trained from scratch. One honest finding: no single model wins everywhere. A version built like an image-spotter (convolutional) is better at pinpointing small, localized things. A version built like a language-reader (transformer) is better at big-picture reasoning. The researchers also showed the models still work well with less training data and cheaper hardware — a big deal for smaller hospitals.

So what? Reading scans is slow, and many hospitals worldwide don't have enough radiologists. A flexible AI that adapts to many jobs and many scanners could mean faster reads, fewer missed findings, and more consistent care outside big-city medical centers. The models are being released inside nnU-Net and nnDetection, two widely used medical imaging toolkits, so hospitals and researchers can try them without rebuilding their software from zero.

The catch: this is a research paper, not an approved medical device. It was tested on curated datasets, not on the messy, real-world mix of patients in an emergency room, and regulators have not cleared it for diagnosis. Think of it as a promising second pair of eyes for a trained radiologist — not a replacement for one, and not something you can ask for at your next scan.