Computational protocol: Human visual cortical responses to specular and matte motion flows

Similar protocols

Protocol publication

[…] Functional imaging was conducted using a Siemens 3T-Trio magnet (Erlangen, Germany) with a 32-channel head coil. To allow participants to have unrestricted viewing of the display through each eye, the coil was operated with the lower 20 elements only. Images were collected with a T*2 sensitive gradient echo imaging pulse sequence (TR = 3 s, TE = 30 ms, delay in TR = 0.8 s, flip angle = 90°, matrix = 96 × 96, GRAPPA acceleration factor = 2, FOV = 192 × 192 mm, partial Fourier=78, voxel size = 2 mm isotropic) in 35 interleaved oblique coronal slices covering the occipital lobes. Stimuli were displayed on an LCD monitor (“BOLDscreen,” Cambridge Research Systems, Kent, UK) with a spatial resolution of 1920 × 1200 pixels, temporal resolution of 60 Hz, and mean luminance of 450 cd/m2. The monitor output was linearized via correction of luminance values measured with a ColorCAL MKII colorimeter (Cambridge Research Systems, Kent, UK). The screen was viewed through a mirror mounted on the head coil at a distance of 112 cm, giving a viewing angle of 26.0° × 16.4°. Stimuli were displayed using PsychToolbox (Brainard, ; Pelli, ; Kleiner et al., ) on a Macbook Pro driving an Intel HD Graphics 4000 video card. As detailed below, analyses were performed using FreeSurfer 5.1.0 (Dale et al., ; Fischl et al., ), FSL 4.1.6 (Smith et al., ), and AFNI/SUMA (2013/05/22; Cox, ; Saad et al., ). [...] The timeseries for each participant and run were first high-pass filtered with Legendre polynomials up to the third degree. An amplitude was then estimated for each block as the mean signal within its five volumes (15 s), shifted by two volumes (6 s) to compensate for the delayed hemodynamic response. The amplitude estimates within each run were then normalized (z-scored). This procedure produced 192 responses per participant for each node on the cortical surface; four responses for each of four conditions in each of 12 runs.The MVPA was performed separately for each participant and visual area, and was implemented using a 12-fold leave-one-run-out strategy in which the responses from a given run were designated (in turn) to form the “test” set and the remaining runs to form the “training” set. In each fold, separate linear support vector machines (SVMs) were trained for image and dot renderings on labeled examples of matte and shiny flow response patterns. Each training set thus consisted of 88 examples, with matte and shiny flow examples equally represented. Flow discrimination accuracy was then estimated by using the trained SVMs to predict the flow condition of test set examples from the same (within-class) or different (between-class) rendering. SVMs were implemented with libsvm 3.17 (Chang and Lin, ) via Matlab (The Mathworks Inc., Natick, MA). The accuracy was based on the proportion of hits and false alarms after aggregation of the 12-folds, and expressed in d′ units. The d′ calculation included an addition of 0.5 to all hit and false alarm counts and an addition of 1 to the number of trials in each condition class, in order to accommodate extreme hit or false alarm rates (Stanislaw and Todorov, ).Features were selected for inclusion in the response pattern for a given visual area based on the t-value of the all stimulus > blank screen contrast performed in the univariate analysis. The surface nodes within a given visual area were ranked in descending order based on the magnitude of this localizing t-value, and the MVPA procedure was performed with patterns formed from including increasing numbers of such ranked nodes (from n = 10 to N, where N is the number of nodes with statistically significant t-values at p < 0.05, one-tailed, uncorrected) in 10 node increments. The variation in MVPA performance with increasing nodes was summarized via a least-squares fit to the function: (1)p=a(1-e-nc) where p is the performance level (d′), n is the number of included nodes, and a and c are fitted parameters that describe the asymptotic performance (a) and curvature (c). The classification accuracy for a given participant, visual area, training class (image, dot), and testing class (within-class, between-class) was then taken as either the fitted asymptote or, in the case of unsuccessful fit, the mean accuracy over nodes. The classification accuracy and fitted performance levels with increasing nodes are shown in Supplementary Figures –. […]

Pipeline specifications

Software tools FreeSurfer, AFNI, LIBSVM
Applications Miscellaneous, Functional magnetic resonance imaging
Organisms Homo sapiens