Radiology AI Foundation Models: Promising But Not Ready

A vision foundation model for radiology is an AI system pretrained on huge volumes of medical scans - brain MRIs, chest X-rays, CT images - so it can later be adapted to many specific tasks like spotting or outlining abnormalities. A new scoping review of 67 studies, published between January 2017 and March 2026, took stock of these models, and its one-sentence conclusion is worth sitting with: they show promising transferability, but clinical translation remains constrained.

What The Review Actually Found

The researchers behind this review mapped radiology-specific vision foundation models across three pillars: how big and varied the training data is, how the architectures and pretraining scale, and how well the models transfer to new tasks.

The datasets leaned heavily toward brain MRI, thoracoabdominal CT, and chest X-ray, and their sizes ranged from fewer than 100,000 samples to multi-million-image cohorts. On the technical side, transformer-based architectures dominated, almost always paired with self-supervised pretraining - techniques like masked image modeling (where the model learns by filling in hidden patches of an image), contrastive learning (learning by telling similar images apart from different ones), and multi-stage approaches that combine methods.

When it came to testing, most studies measured segmentation (tracing structures pixel by pixel) and classification (labeling a scan as one thing or another). That's where the pattern gets interesting.

The Gap Between Lab Scores And Clinical Reality

Here's the part most coverage glosses over: a model can ace segmentation benchmarks in one hospital's data and still fail quietly somewhere else. The review found that validation under real-world shifts - cross-center differences, cross-scanner differences, changes in anatomy or imaging modality - was inconsistently reported. In other words, we often don't know whether these models generalize beyond the exact conditions they were trained in.

It's easy to conflate "good benchmark performance" with "ready for the clinic," but they're not the same thing. The review names four specific constraints: limited data representativeness (the training scans don't reflect the full diversity of real patients and machines), heterogeneous benchmarks (every study measures differently), incomplete reporting (you can't audit what wasn't written down), and insufficient deployment-oriented evaluation (almost nobody tests these models the way they'd actually run in a hospital workflow).

What This Actually Points To

The real story here isn't that radiology foundation models are overhyped - it's that the field is measuring them with tools built for a different question. "Can this model segment tumors well on this dataset?" is a solvable research question. "Will this model behave safely on a scanner from a different vendor at a hospital it's never seen?" is the clinical question, and per this review, it's rarely being asked.

My read as an interpretation: the promising transferability results are genuinely encouraging - self-supervised pretraining on large radiological cohorts does seem to produce models worth adapting. But until shift-testing becomes standard reporting rather than an optional extra, every benchmark number should come with an invisible asterisk. The bottleneck isn't architecture anymore; it's evidence.