Visual Image Reconstruction from fMRI Activation Using Multi-scale Support Vector Machine Decoders

摘要

The correspondence between the detailed contents of a person's mental state and human neuroimaging has yet to be fully explored. Previous research reconstructed contrast-defined images using combination of multi-scale local image decoders, where contrast for local image bases was predicted from fMRI activity by sparse logistic regression (SLR). The present study extends this research to probe into accurate and effective reconstruction of images from fMRI. First, support vector machine (SVM) was employed to model the relationship between contrast of local image and fMRI; second, additional 3-pixel image bases were considered. Reconstruction results demonstrated that the time consumption in modeling the local image decoder was reduced to 1% by SVM compared to SLR. Our method also improved the spatial correlation between the stimulus and reconstructed image. This finding indicated that our method could read out what a subject was viewing and reconstruct simple images from brain activity at a high speed.