Neuroscience research adapts visual brain decoding to reconstruct imagined content from fMRI data

A recent study in neuroscience reveals that visual brain decoding technology can be modified to recreate imagined content using fMRI data. Researchers introduced a latent functional alignment method that aligns brain activity triggered by imagery into a pretrained model's conditioning space, while other elements remain unchanged. This approach was evaluated with the Imagery-NSD benchmark, which is based on the Natural Scenes Dataset (NSD). To tackle the challenge of limited matched imagery-perception supervision, the team implemented a retrieval-based augmentation strategy that selects semantically similar NSD perception trials. The research, documented as arXiv:2604.15374v1, highlights interdisciplinary collaboration between neuroscience and artificial intelligence. Four human participants were involved, and results indicated significant enhancements in high-level semantic reconstruction using the latent functional alignment method. Current visual brain decoding frameworks have mainly focused on perception tasks, leaving their efficacy in mental imagery less explored. The study utilized extensive neuroimaging datasets and advanced diffusion-based generative models, with the DynaDiff model, a leading perception decoder, forming the basis for adapting to imagined content reconstruction.