A continually expandable foundation model for brain MRI

2026-08-08

AI Summary

This work proposes Alcmaeon, a continuously extensible 3D foundation model for brain MRI that addresses the limited generalizability and catastrophic forgetting prevalent in existing models, which are often constrained to specific diseases, populations, or imaging protocols. Trained unsupervised on over 425,000 unlabeled brain MRI scans, Alcmaeon incorporates a novel Graph-Blueprint Pruning mechanism that dynamically preserves critical network modules when integrating new clinical domains, thereby mitigating catastrophic forgetting. The model demonstrates consistently lower voxel-level reconstruction forgetting across diverse applications—from healthy aging to tumor imaging—and its hierarchical representations effectively support multiple downstream tasks, including image synthesis, disease classification, survival modeling, and postoperative outcome prediction, while providing interpretable records of capacity allocation.

Abstract

Brain magnetic resonance imaging (MRI) is central to neuroscience and clinical assessment, but models are commonly developed for individual diseases, populations or imaging protocols. Foundation models promise more general representations, yet they are usually pretrained once and can lose earlier capabilities when updated with new data. Here we show that Alcmaeon, a three-dimensional brain MRI foundation model pretrained without manual labels on more than 425,000 volumes and derived imaging maps, can be expanded sequentially across clinical domains. Alcmaeon combines volumetric encoding and latent diffusion generation with Graph-Blueprint Pruning (GBP), which protects network modules important to earlier domains while leaving the remaining capacity trainable. Across expansion from healthy ageing and neurodegeneration to developmental, psychiatric and tumour imaging, GBP showed less forgetting than sequential adaptation and elastic weight consolidation across voxel-level reconstruction measures, with its largest advantage after adaptation to tumour imaging. The blueprints provided an inspectable record of how model capacity was protected and reused. Representations from different model levels supported image synthesis, disease classification, survival modelling and postoperative prediction, although no single representation was optimal for every task. These findings provide a route towards brain MRI foundation models that can grow with emerging data while retaining earlier capabilities.

Problem

• foundation model
• brain MRI
• catastrophic forgetting
• continual learning
• model expansion

Innovation

• foundation model
• continual learning
• Graph-Blueprint Pruning
• brain MRI
• catastrophic forgetting