Dublin City University Trains MRI to Spot Alzheimer's Without Expensive PET Scans
A BiomedCLIP-based teacher learns from paired PET-MRI scans, then teaches an MRI-only student to flag amyloid buildup without expensive PET imaging.
ProTakeaways
- New framework predicts amyloid-beta positivity from MRI alone, no PET or clinical variables needed at inference.
- Best AUC of 0.74 on OASIS-3 and 0.68 on ADNI across four MRI contrasts.
- Improves T1w-only AUC by 19.7% over prior MRI baseline (Kim et al., 0.61 to 0.73).
- Uses BiomedCLIP ViT with LoRA, cross-modal attention teacher, and Centiloid-aware triplet mining.
- Code released at GitHub repo; paper on arXiv.
- Saliency maps confirm the model focuses on anatomically plausible cortical regions rather than artifacts.
Detecting the amyloid-beta plaques that mark early Alzheimer's disease usually means booking a PET scan, which is expensive, radioactive, and unavailable in most clinics. A team from Dublin City University has now shown that a knowledge distillation setup can teach a standard MRI model to predict amyloid positivity on its own, hitting an AUC of 0.74 on OASIS-3 and 0.68 on ADNI without any PET or clinical variables at inference time.
Why PET has become a bottleneck
Amyloid PET is costly, invasive, and thin on the ground, which rules it out for population-level screening. With FDA approvals of lecanemab and donanemab (both of which require confirmed amyloid pathology before treatment), projected demand for amyloid PET is growing roughly 20-fold.
MRI is the obvious alternative: non-invasive, cheap, and already part of routine dementia workups. The catch is that MRI can't directly image amyloid, and prior MRI-only models have been mediocre. A 3D EfficientNet baseline from Kim et al. hit just AUC 0.61 on T1w scans. Models that do better tend to lean on non-imaging covariates like APOE genotype or cognitive scores, but those are missing in a huge fraction of real records. Routine APOE testing isn't recommended by ACMG guidelines and is absent for 26% of NACC participants, and cognitive measures are missing from 89% of EHR dementia records.
The three-phase teacher-student setup
The core idea is cross-modal knowledge distillation. A BiomedCLIP-based teacher learns PET-MRI alignment through cross-modal attention and triplet contrastive learning with PET-informed (Centiloid-aware) online negative sampling. An MRI-only student then mimics the teacher through feature-level and logit-level distillation. BiomedCLIP is a Vision Transformer pretrained on 15 million biomedical image-text pairs, which the authors adapt using LoRA on the last six transformer blocks.