Brain MRI Learns Not to Forget express
A doctor spends years learning to recognize diseases from brain scans. First, they excel at spotting signs of aging, then they master tumors, and suddenly they start confusing what they knew before. Many neural networks behave much the same way: new knowledge overwrites the old.
The new model Alcmaeon tackles this problem with an approach reminiscent of margin notes in a textbook. The Graph-Blueprint Pruning method identifies "anchor summaries" in the network — groups of connections critical for already learned diseases. These summaries are frozen, while the rest of the network freely learns from new data. The model was tested on four areas: healthy aging, neurodegeneration, brain development, and tumors. When transitioning to tumors, old skills were preserved best — the network did not unlearn what it had seen before.
Interestingly, the "summaries" turned out to be different for different tasks. For disease classification, certain network layers worked better; for survival prediction, others. No universal recipe was found, but the approach itself offers hope for creating truly durable medical assistants. After all, unlike a computer, the brain can learn throughout life without erasing old memories, and Alcmaeon takes a step in that direction.
Original abstract
as written by the authors arXiv:2608.08319v1
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.
The Graph-Blueprint Pruning method works like "protected notes" in a notebook: you can't accidentally erase important formulas, but on blank pages you can write anything.