Scientists have developed an artificial intelligence system capable of reconstructing images a person is viewing by analyzing their brain activity, marking a significant advance in the field of neuroimaging and machine learning. The system, named Brain-IT, processes functional magnetic resonance imaging (fMRI) data— which measures blood flow and oxygenation changes in the brain— to generate images that closely replicate what the subject is looking at, ranging from faces and objects to complex scenes.

Leading the research is Professor Michal Irani of the Weizmann Institute of Science in Israel, who highlighted Brain-IT’s potential to eventually extend beyond viewed images to interpreting dreams. Unlike earlier models that required extensive, individualized training with participants spending numerous hours inside MRI machines, Brain-IT can adapt to a new person’s brain activity in approximately one hour. This improvement addresses a major limitation of prior systems, which often needed up to dozens of hours of scanning per participant under costly, claustrophobic conditions.

The team’s approach involved two interconnected AI models operating in opposite directions: one converts brain activity into an image, while the other predicts brain scan patterns based on visual stimuli. By iteratively translating images back and forth between these models, the researchers were able to generate vast amounts of synthetic training data, circumventing the typical scarcity of high-quality fMRI datasets. This process allowed the AI to effectively “build” a sizable dataset without requiring additional costly brain scans.

Brain-IT’s architecture also leverages the identification of 128 distinct brain regions that appear to function similarly across individuals. Some of these areas align with previously known neuroscience findings, while others represent novel discoveries. According to the researchers, the system excels not only in capturing the content of viewed images but also in preserving their finer details, such as color— a challenge that earlier models struggled to meet.

The team emphasized that Brain-IT’s improved speed and accuracy stem from its ability to recognize recurring patterns of neural activation across different people as they process visual information. While still in the research phase, the system’s development offers promising avenues for future applications in neuroscience, cognitive science, and potentially clinical diagnostics.