A continually expandable foundation model for brain MRI
The paper introduces Alcmaeon, a continuously expandable 3D brain MRI foundation model trained on over 425,000 unlabeled volumes that utilizes Graph-Blueprint Pruning to sequentially adapt to diverse clinical domains while effectively preventing catastrophic forgetting and preserving earlier capabilities.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are trying to teach a robot to understand the human brain. You show it thousands of MRI scans—those detailed 3D pictures doctors use to look inside our heads. At first, the robot learns to recognize healthy brains. But then, you want it to spot brain tumors, or understand the changes caused by Alzheimer's, or even see how a child's brain is still growing. The problem is, when you teach the robot something new, it often forgets what it learned before. It's like if you studied for a history test, then immediately started studying for a math test, and suddenly you couldn't remember the dates of the French Revolution. In the world of artificial intelligence, this is called "catastrophic forgetting." Scientists have been trying to build "foundation models"—super-smart AI brains that learn general rules from massive amounts of data so they can be adapted to many different tasks without needing to be rebuilt from scratch every time. But making these models work for 3D brain scans is incredibly hard because the data is huge, messy, and comes from many different types of machines and patients. The big question is: Can we build a brain-AI that keeps learning new things without losing its old memories?
Enter Alcmaeon, a new kind of AI foundation model for brain MRI scans that acts like a very disciplined student who never forgets. Created by a team of researchers, Alcmaeon was trained on over 425,000 three-dimensional brain scans without anyone having to manually label them. Instead of just learning once and stopping, Alcmaeon is designed to grow. It learns sequentially, moving from healthy populations to neurodegenerative diseases, then to developmental and psychiatric conditions, and finally to brain tumors. The secret sauce that makes this possible is a clever trick called Graph-Blueprint Pruning (GBP).
Think of Alcmaeon's brain as a massive library filled with millions of books (or in this case, tiny computer circuits called "modules"). When the AI learns a new topic, like spotting brain tumors, it usually tries to rewrite its entire library, which accidentally erases the old books about healthy brains. GBP is like a librarian who puts a "Do Not Touch" sticker on the specific books that are essential for the old topics. When the AI learns something new, it can only write in the empty spaces or use the books that don't have stickers. This way, the AI can learn about tumors without accidentally deleting its knowledge of healthy aging. The "blueprint" is essentially a map that keeps track of which books are protected and which are free to be updated.
The researchers tested this by teaching Alcmaeon through four distinct stages: healthy aging, neurodegeneration, developmental/psychiatric imaging, and finally, brain tumors. They compared Alcmaeon using this "blueprint" method against other common ways of teaching AI, like simply retraining everything (which causes forgetting) or using a method called Elastic Weight Consolidation (EWC). The results showed that when the AI faced the biggest challenge—learning about brain tumors, which look very different from healthy brains—Alcmaeon with GBP kept its memory much better than the others. While the other methods started to forget how to reconstruct healthy brain images (dropping their performance scores significantly), Alcmaeon with GBP held onto its skills, keeping its ability to see healthy brains intact while still learning to spot tumors.
Interestingly, the paper found that making the AI model bigger (adding more parameters) didn't automatically make it better at remembering things. Instead, how the model was built mattered more. For instance, the type of mathematical "language" the model used to understand the images (deterministic vs. variational) changed how well it worked for different types of scans. For standard structural scans, a straightforward approach worked best, but for some complex micro-structure maps, a more flexible approach was needed.
The team also checked if this "remembering" AI could actually help doctors. They tested if the knowledge Alcmaeon gathered could be used for real-world tasks like predicting how long a patient with a brain tumor might live, classifying Alzheimer's disease, or even generating new MRI images to see what a brain might look like without a tumor. They found that different parts of the AI's brain were good at different jobs. The "middle layers" of the model were great at spotting diseases, while the whole system was good at generating new images. Crucially, even after learning about tumors, the parts of the AI that knew about Alzheimer's and other diseases were still there and useful, proving that the "Do Not Touch" stickers worked.
In short, this paper suggests that we don't need to build a new AI for every single brain disease. Instead, we can build one expandable AI that learns step-by-step, protecting its past knowledge while absorbing new information. While the researchers are careful to say this is a promising step forward and not a perfect solution yet (and that it needs more testing on real patients before it can be used in hospitals), they have shown a clear path toward AI that can grow with medical science without losing its way.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.