New AI Tool Lets Brain Scans Answer Questions Without Retraining

Brain scans like fMRI produce complex data that's hard to analyze. Researchers have created many AI models, each good at a different aspect of brain data. But combining them has been difficult because each model uses different formats. Now, scientists have developed a modular approach: each AI model becomes a separate tool that a central AI agent can use as needed. In tests with 76 questions, this method improved the overall quality of answers by 0.70 points on a 5-point scale.

The key innovation is that no retraining is required. Just like using different apps on your phone without rewriting them, the central AI can call upon specialized brain models alongside classical analysis tools and literature. This makes it much easier to leverage the strengths of multiple models. For example, one model might be good at identifying brain regions, while another excels at detecting patterns across the whole brain. By using both, the AI agent can provide more accurate and comprehensive answers.

The study found that the improvement mainly comes from how the AI agent selects and combines information from different tools. Interestingly, the two brain models tested won on different types of questions, but using both together matched the performance of the better one. However, one model had a higher failure rate due to running out of context space. This modular approach could be extended to other types of brain data like EEG and MRI, potentially accelerating research into brain disorders and treatments.