The contrast sensitivity function (CSF) characterizes visual function, and is widely used in research on visual perception and ophthalmological disorders. The CSF describes the lowest contrast level that participants can perceive as a function of spatial frequency. Here, we present a new method to estimate the neural equivalent of the CSF that describes how a population of neurons responds to contrast as a function of spatial frequency. Using functional magnetic resonance imaging (fMRI) at 7 Tesla, we measured neural responses while participants viewed gratings that varied systematically in contrast and spatial frequency. We modeled the neural CSF (nCSF) using an asymmetric parabolic function, and we modeled the transition from no response to full response using a contrast response function (CRF). We estimated the nCSF parameters for every cortical location by minimizing the residual variance between the model predictions and the fMRI data. We validated the method using simulations and parameter recovery. We show that our nCSF model explains a significant amount of the variance in the fMRI time series. Moreover, the properties of the nCSF vary according to known systematic differences across the visual cortex. Specifically, the peak spatial frequency that a cortical location responds to decreases with eccentricity and across the visual hierarchy. This new method will provide valuable insights into the properties of the visual cortex and how they are altered in both healthy and clinical conditions.

Much of the visual system is organized into visual field maps. In humans, this organization can be studied non-invasively by estimating the receptive fields of populations of neurons (population receptive fields; pRFs) with functional magnetic resonance imaging (fMRI). However, fMRI cannot capture the temporal dynamics of visual processing that operate on a millisecond scale. Magnetoencephalography (MEG) does provide this temporal resolution but generally lacks the required spatial resolution. Here, we introduce a forward modeling approach that combines fMRI and MEG, enabling us to estimate pRFs with millisecond resolution. Using fMRI, we estimated the participant’s pRFs using conventional pRF-modeling. We then combined the pRF models with a forward model that transforms the cortical responses to the MEG sensors. This enabled us to predict event-related field responses measured with MEG while the participants viewed brief (100 ms) contrast-defined bar and circle shapes. We computed the goodness of fit between the predicted and measured MEG responses across time using cross-validated variance explained. We found that the fMRI-estimated pRFs explained up to 91% of the variance in individual MEG sensor’s responses. The variance explained varied over time and peaked between 75 ms to 250 ms after stimulus onset. Perturbing the pRF positions decreased the explained variance, suggesting that the pRFs were driving the MEG responses. In conclusion, pRF models can predict event-related MEG responses, enabling routine investigation of the spatiotemporal dynamics of human pRFs with millisecond resolution.