Towards Brain MRI Foundation Models for the Clinic: Findings from the FOMO25 Challenge
The FOMO25 challenge demonstrates that self-supervised foundation models trained on large, heterogeneous clinical datasets can outperform supervised in-domain baselines in few-shot and out-of-domain brain MRI tasks, revealing that optimal pretraining objectives vary by task and that performance gains do not strictly depend on scaling model size or training duration.
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 be a doctor. Specifically, you want it to look at brain scans (MRI images) and spot problems like strokes, tumors, or just tell you how "old" a brain looks.
In the past, to teach this robot, we had to show it thousands of scans where a human expert had already drawn circles around the problems. This is like hiring a team of art teachers to spend years drawing on every single picture before the student can learn. It's expensive, slow, and we run out of pictures with drawings very quickly.
The Big Idea: The "Self-Taught" Student
The paper describes a challenge called FOMO25 (Foundation Model Challenge for Brain MRI 2025). The goal was to see if we could teach the robot to learn on its own first, using millions of brain scans that don't have any drawings or labels.
Think of it like this: Instead of giving the student a textbook with answers, we give them a library of 60,000 blank brain scans. We tell them, "Look at these pictures, figure out what a brain looks like, how it's structured, and what normal variations are." This is called Self-Supervised Learning. The robot learns the "language" of brains by trying to reconstruct missing parts of the images or by comparing different scans, without needing a human to say "this is a tumor."
Once the robot has studied this massive library, we give it a tiny test: "Here are 20 scans with a few labeled tumors. Can you now find the rest?" This is called Few-Shot Learning.
The Challenge Setup
The organizers set up a race with two tracks:
- The Method Track: Teams could only use the 60,000 scans provided by the organizers.
- The Open Track: Teams could use any data they wanted, including private hospital data or other massive datasets.
They tested the robots on three real-world tasks:
- Spotting Strokes: Is there a damaged area? (Classification)
- Mapping Tumors: Where exactly is the tumor? (Segmentation)
- Estimating Age: How old does this brain look? (Regression)
Crucially, they tested these robots on data from different hospitals and countries than where they were trained. This is like teaching a student in Copenhagen and then testing them on patients in India. Real life is messy, and scans look different depending on the machine used.
The Surprising Results
Here is what the race revealed, translated into simple terms:
The "Self-Taught" Students Won:
The robots that did the "self-study" (pretraining) on the huge unlabeled library were much better at handling the messy, real-world data than robots trained only on the small, labeled test sets. In fact, the best "self-taught" robots were even better than the "specialist" robots trained specifically on the exact same type of data they were being tested on.- Analogy: It's like a student who read every book in the library being better at a specific trivia question than a student who only memorized the answer key for that one question.
One Size Does Not Fit All:
There wasn't one single "magic trick" for learning that worked for everything.- If you wanted the robot to find tumors (segmentation), a learning method that focused on filling in missing puzzle pieces (called MAE) worked best.
- If you wanted the robot to spot strokes (classification), a method that compared different images to find similarities (hybrid methods) worked better.
- Analogy: It's like using a hammer to drive a nail and a screwdriver to turn a screw. You need the right tool for the specific job.
Bigger Isn't Always Better:
In the world of AI, people often think "bigger model = smarter." But here, making the robot's brain (the model) huge or making it study for longer didn't necessarily make it perform better on the final test.- Analogy: It's like a student who reads a million books but doesn't know how to study for the specific test. A smaller, well-trained student sometimes outperforms the giant one.
Small Models Can Be Champions:
Some of the winning teams used relatively small models. This is great news because it means we don't need supercomputers the size of a city to build these medical tools; we can build them on smaller, more accessible hardware.
Why This Matters for the Clinic
The biggest takeaway is that Foundation Models are ready for the real world.
For a long time, AI in medicine was great at solving clean, perfect puzzles in a lab but failed when faced with the messy reality of a hospital (different machines, different patients, different countries). This challenge proved that if we let AI learn from the vast, unlabeled data that hospitals already have, we can build robust tools that work even when the data is imperfect.
In a Nutshell:
The FOMO25 challenge showed that we can train AI doctors by letting them "read" millions of unlabeled brain scans first. This makes them smart enough to handle the messy reality of real hospitals, often beating traditional methods, without needing massive amounts of expensive human labels. It's a giant step toward having AI that can actually help doctors in everyday practice, not just in research papers.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.