BrainGFM uses synthetic scans to stretch scarce brain-imaging data

Brain-imaging AI faces a practical mismatch: researchers want systems that can outline tumors, distinguish disease categories and improve noisy scans, but each task needs suitable training data. Those data can be scarce, especially when experts must carefully label anatomical structures or lesions.

A preprint by Yuting He, Xu Wen and colleagues presents BrainGFM, a generative foundation model designed to make previously learned brain-imaging knowledge reusable across these different jobs. Rather than only supplying pretrained features to another model, it can generate new training examples or undergo targeted adaptation for image reconstruction.

The researchers report evaluations across 48 benchmark tasks, followed by clinical analyses using stroke cohorts from six hospitals. Their findings support the approach’s versatility, while also exposing important differences between convincing synthetic images and dependable clinical assistance.

A shared foundation, two ways to use it

The team assembled a collection of 1,295,080 two-dimensional slices from 64,754 MRI and CT volumes drawn from 11 resources. It covered CT and several MRI sequences, including T1, T2, contrast-enhanced T1, FLAIR and diffusion-weighted imaging, or DWI. About 144,000 slices were held out; approximately 1.15 million formed the pretraining partition.

BrainGFM represents images and segmentation masks—the outlines used to label regions—as discrete visual tokens. A transformer learns to generate these representations from coarse spatial scales to finer ones. The authors describe the resulting learned distribution as a reusable brain-imaging “prior”: knowledge of the patterns represented in its training collection.

For segmentation and classification, the model generates task-specific synthetic data that train conventional prediction systems. For reconstruction, it instead learns from paired source and target images. Task adaptation uses small, trainable additions to the transformer while keeping its original weights and visual tokenizer frozen.

Synthetic data helped low-data learning

Across 20 segmentation benchmarks, adaptation generally used 20 annotated slices from one target case. BrainGFM then produced 2,000 synthetic image-mask pairs per task to train a U-Net segmentation model.

The resulting models achieved a mean Dice score of 72.4%, compared with 54.4% for direct low-data U-Net training and 70.0% for MedSAM. Dice measures agreement between predicted and reference regions. BrainGFM also exceeded the 71.4% mean of the study’s full-data U-Net reference, although that comparison concerns the particular datasets and training protocols tested.

For eight classification tasks, the researchers used one case per class in volumetric settings and generated 1,000 images per class. Mean AUC, which measures discrimination across decision thresholds rather than simple accuracy, reached 94.0%. That compared with 89.4% for direct low-data training and 94.5% for the full-data reference.

Twenty further benchmarks examined image enhancement or transformation. Importantly, the low-field MRI and low-dose CT inputs were created through simulated degradation, rather than acquired directly under those conditions. Other tasks predicted contrast-enhanced MRI appearances from non-contrast scans.

Stroke studies tested clinical relevance

The clinical dataset comprised 4,796 patients from two development hospitals and four external hospitals. Eligible subsets supported lesion segmentation, five-category stroke subtyping and generation of FLAIR images from DWI. Low-data segmentation and subtyping used 60 real training cases, rather than the one-case benchmark setting.

On external cohorts, BrainGFM-supported lesion segmentation reached a Dice score of 76.9%, and stroke subtyping achieved an AUC of 84.5%.

Three radiologists also rated acquired and generated FLAIR images from 60 stroke cases without knowing their source. BrainGFM’s mean overall diagnostic-quality score was 3.51 out of five, close to 3.57 for acquired FLAIR. But only 15.6% of its assessments scored at least four on every quality dimension, versus 51.1% for acquired images. Similar average ratings therefore did not establish interchangeability.

Assistance was useful, but uneven

In a separate 300-patient study, three radiologists assessed three stroke subtypes first without, then with, BrainGFM predictions. Mean accuracy rose from 55.6% to 63.4%: a 7.8-percentage-point gain, or 14% relative improvement. Standalone AI accuracy was 79.7%.

The average concealed substantial variation. One reader improved from 53.0% to 75.0%, another declined from 57.3% to 56.0%, and the third improved modestly. Incorrect AI recommendations sometimes overturned initially correct judgments.

The authors identify predominantly two-dimensional modeling and limited population, disease and acquisition coverage as constraints. They also call for prospective clinical validation. BrainGFM’s central promise is more flexible reuse of imaging data—not proof that synthetic scans can replace measurements or that AI advice consistently improves every clinician’s decisions.