A cortical region consisting entirely of face-selective cells

Abstract

Face perception is a skill crucial to primates. In both humans and macaque monkeys, functional magnetic resonance imaging (fMRI) reveals a system of cortical regions that show increased blood flow when the subject views images of faces, compared with images of objects. However, the stimulus selectivity of single neurons within these fMRI-identified regions has not been studied. We used fMRI to identify and target the largest face-selective region in two macaques for single-unit recording. Almost all (97%) of the visually responsive neurons in this region were strongly face selective, indicating that a dedicated cortical area exists to support face processing in the macaque.

Fig. 1. Targeting an fMRI-identified face patch for single-unit recording

(A) A semi-sagittal section through the right hemisphere of monkey M1 showing three face-selective patches along the STS. Single-unit recordings in monkeys M1 and M2 were targeted to the middle face patch, located ~6 mm anterior to the interaural line; the red rectangle indicates a coronal slice passing through the middle face patch. The three white arrows point to the lesion left by the recording guide tube. (B) Two coronal slices showing the middle face patch in monkeys M1 (left) and M2 (right) (at A6.5 and A5.5, respectively). MION activation is overlaid on raw functional echo planar (EPI) images. Arrows point to the specific region targeted for electrophysiology in each monkey. In monkey M1, the targeted face patch was located on the lower lip of the STS in the right hemisphere. In monkey M2, the targeted face patch was located in the fundus of the STS in the left hemisphere. (C) Face patches and guide-tube track in three dimensions, rotated into the coordinate system of the recording grid (monkey M1). After chamber implantation, a high-resolution anatomical scan was obtained with six oil-filled markers positioned inside a grid in the recording chamber. We determined which grid hole to use by rotating the brain, together with registered face-selective fMRI activation, into the coordinate system defined by these markers. This panel shows three orthogonal slices passing through the point marked by the intersection of the red lines. Cells in this monkey were recorded from the hole at the intersection of the red lines, and from two adjacent, more-medial holes in the same row of the grid. The dark elongated lesion confirms that the guide tube passed through this point to accurately and precisely target the middle face patch. We recorded from all cells encountered between the start of the gray matter and the start of the white matter in the lower bank of the STS. Scale bar, 1 cm. P < 10⁻⁴ for MION activations.

Fig. 2. Face selectivity of single units in the middle face patch

(A) Selectivity profiles of all visually responsive cells in monkeys M1 (left) (182 cells) and M2 (right) (138 cells) to 96 images of faces, bodies, fruits, gadgets, hands, and scrambled patterns (16 images per category, see fig. S1 for stimuli). Each row represents one cell and each column one image. The rows were sorted by the FSI, and the columns were sorted by image category. To compute selectivity profiles for each cell, responses to the 96 images were averaged over a 200-ms interval starting at the response latency, the baseline (the average response from 0 to 50 ms) was subtracted, and the response normalized. The average response time course to each of the 96 images is shown in fig. S6. (B) Average response to each of the 96 images across all visually responsive cells in monkeys M1 and M2. Error bars represent ±1 SE. The black line indicates six average SEs. (C) Nonface images that elicited a response above six average SEs in monkeys M1 and M2. Images are sorted from left to right by decreasing elicited response magnitude. (D) Distribution of FSIs across all visually responsive cells. Dotted lines indicate |FSI| = 0.33 (corresponding to a 2:1 ratio of face-to-nonface object response).

Fig. 3. Face identity and category information are carried by the population of face-selective cells

(A) Matrix of Euclidean distances between each of the 96 × 96 pairs of test and template response vectors. The maximum possible distance between a test and a template vector is $\sqrt{94}$. (B) The top row shows the percent correct identification and categorization for each image, based on a nearest neighbor algorithm. The middle row shows the same data grouped by category (F, faces; B, bodies; Fr, fruits; G, gadgets; H, hands; S, scrambled patterns). Error bars represent ±1 SE. The bottom graphs are the percentages of correct identification and categorization for six different categories as a function of time after stimulus presentation, computed using a 50-ms sliding window. Chance performance would be 1/96 for identification and 1/6 for categorization (indicated by the horizontal line in each graph). Data are from monkey M1.