Abstract
Introduction: A recent focus of fMRI analysis is to predict features of the presented stimulus from measured brain signals. Most of these methods use machine-learning techniques, though some use biological models. Biological approaches explicitly model the underlying neural mechanism (Kay et al, 2008) and are therefore independent of predefined stimulus categories. Using a biological model based on population receptive field (pRF) properties with minimal parameters, we show that it is possible to identify any visual image.
Methods: We measured fMRI responses elicited by different visual stimuli using 7T MRI. The stimuli consisted of conventional bar-shaped mapping stimuli, semi-random synthetic stimuli and natural images. First, we estimated the population receptive field properties of each cortical location (Dumoulin and Wandell, 2008). Next, we identified the presented image based on measured fMRI signals and the pRF model. Both steps used distinct stimuli. Using visual field maps V1 to V3, the pRF model predicted fMRI signal amplitudes for many different images. We correlated the measured fMRI amplitudes with those predicted from the pRF model. The presented image was identified by choosing the pRF model prediction with the highest correlation.
Results: Image identification from a dataset of 1000 images is far above chance for visual field maps V1 to V3, but decreases in performance (V1:~64%, V2:~37%, V3:~32%, chance=0.1%, synthetic images). Both standard mapping stimuli and semi-random synthetic stimuli can be used to estimate the pRF properties. Performance is slightly better when the pRF model is trained on similar synthetic images as compared to standard mapping stimuli (V1: 67% versus 60%, respectively). Identification of natural images is also possible, but less accurate.
Discussion: The pRF method is a fast and simple biological model, and even when training it on standard mapping stimuli, it can be used for identification of any visual image, including natural images.
Meeting abstract presented at VSS 2012