A new AI framework treats brain PET scans the way clinicians do - by focusing on what the metabolism is actually doing, not just what the image looks like.
Researchers introduced ReMAP-PET, a framework built on a partially-tuned MedicalNet 3D ResNet-50 encoder trained on 1,015 paired PET and standardized uptake value ratio (SUVR) samples. Rather than processing PET as generic 3D volume data - the approach most existing brain foundation models take - it supervises the encoder using regional SUVR profiles through joint regression and contrastive objectives. The result: a 0.070 SUVR mean absolute error and 77.8% PET SUVR Recall@1, beating five frozen pretrained baselines on both counts. The team also linked the metabolic embeddings to clinical language by aligning them with frozen BioClinicalBERT, enabling end-to-end report generation from PET input.
Most medical imaging AI treats modality differences as a nuisance to abstract away. PET is not structural MRI - it captures glucose metabolism and is central to diagnosing Alzheimer's, Parkinson's, and other neurodegenerative conditions. A model that understands regional metabolic patterns, rather than raw voxel intensities, produces embeddings that hold up on diagnostic classification and cognitive regression tasks without any task-specific fine-tuning - a meaningful bar for clinical utility.
The work is a preprint and has not yet been peer-reviewed, so claims about clinical readiness should be read accordingly. That said, the direction - grounding imaging encoders in the structured biological information specific to a modality - is one the field has been slow to pursue, and the linear-probing results suggest the embeddings carry real signal rather than learned shortcuts.