Fig 1.
Examples of the pRF mapping stimuli used to estimate the pRF properties.
The stimuli were either filled with a binarized bandpass filtered noise pattern (A) or natural image content (B).
Fig 2.
(A) The synthetic stimuli were made using a grid of 60 hexagons (cells) that could be filled with either binarized bandpass-filtered noise or mean-luminance gray. Examples of the stimuli used in the experiment; (B) a non-random synthetic stimulus, (C) random synthetic stimuli and (D) natural images.
Fig 3.
Schematic diagram of the image identification pipeline.
A: First, we estimated the parameters of the pRF for every voxel based on the pRF bar stimuli. The pRF is modeled as a circular symmetric Gaussian function of which we use the parameters for position and size. B: Second, we predicted the response profiles for a large set of candidate images by either summing the overlap of the stimulus with each voxel’s pRF (for the synthetic images), or by calculating the RMS-contrast inside each voxel’s pRF (for the natural images). C: Finally, we predicted which candidate image elicited a measured response profile by finding which candidate image’s predicted response profile was most strongly correlated to this measured response profile (i.e. which had the highest pearson’s r).
Fig 4.
The correlation matrix for an example set of synthetic images.
The correlation matrix shows the prediction accuracy using the voxels of V1 from the image identification process for an example set of the synthetic images. The colors represent the correlation (Pearson’s r) of the measured response profiles from all the images with their predicted response profile (from the pRF-model). For this image set, 14 out of 15 images were identified correctly, giving a prediction accuracy of 93.3%.
Fig 5.
Image identification performance both for synthetic (A) and natural image stimuli (B, C).
The pRFs were estimated in a separate scan where the bars were filled with synthetic (A, C) and natural (B) stimuli either in the same session (A,B) or in different sessions on different days (C). The blue, red and green lines show the performance for visual field maps V1, V2 and V3 respectively, for 2 subjects (closed and dashed colored lines). The thickness of the lines includes the 95% confidence intervals. For the synthetic image stimuli we increased the set size up to 1000 different images, for the natural images we increased the set size up to 200 different images. The black dashed line indicates the chance level. The high-contrast synthetic images were most accurately identified (A), and the natural images were also identified far above chance for all candidate image set sizes (B). Furthermore, the identification of the natural images is also possible with the pRF-model that was estimated using standard bar stimuli containing moving checkerboards, and estimated from a separate scanning session on a separate day (C).
Fig 6.
The identification confidence per natural image for our two subjects.
Every dot represents an individual image. We see similar confidence of the individual natural images across the two subjects. Some images are identified less accurately using the pRF-model than others. This is explained by two factors: (i) certain images are more similar in terms of their contrast-energy content and (ii) responses to these images depend more on features that are not captured by the contrast-energy pRF-model predictions.