3.5 Propiedades físicas y mecánicas de la roca
3.5.2 Comportamiento Mecánico de las Discontinuidades
The methods described in the section above propose the means of avoiding the issue of overlapping training and test sets in a manner which, when using a sufficiently large exclusion window, should eliminate the regions where the extended nature of the haemodynamic
response may cause overlapping information. A number of analyses were performed using data collected for a study on the effects of learning on decision templates in the visual cortex carried out by Kuai et al (2013). In this section a brief overview of the data collected and pre- processing performed by Kuai et al (2013) is presented, followed by the MVPA and control analyses performed to evaluate the validity of the leave-c-out cross-validation method and determine the size of exclusion window necessary to eliminate overlap.
3.3.1 Observers
Nine healthy students (age=21.1+-0.75) from the University of Birmingham volunteered for the study. They participated in fMRI scans both before and after a set of behavioural training sessions, and seven of the participants took part in a further post-training scan involving stimuli which differed in size from those they had trained with.
3.3.2 Stimuli
The stimuli presented were pentagons formed from 30 Gaussian dots. Two classes of stimuli were generated by varying the length of the sides of the pentagons. Multiple levels of stimuli between Class I and Class II were generated by use of linear morphing.
3.3.3 fMRI Design
All fMRI sessions comprised seven or eight runs which used an event related design. Seven conditions were presented in each run; six were stimulus conditions with stimuli levels taken from the range provided by the linear morphing, with the seventh being a fixation condition consisting of a fixation point at the centre of the screen. In each run 14 trials per condition were presented. In addition to the 98 trials of fixation and stimuli conditions, an additional three 9s blocks of fixation were included at the beginning, middle and end of each run.
Trial presentations were designed to align with the volumes acquired with a duration of 2 volumes (3s). The first volume covered the 200 ms stimulus presentation and a 1300 ms blank. The second volume of each trial covered the participant's response; a colour cue was presented for 1200 ms followed by a fixation for 300 ms. During this time participants were expected to press a key indicating the class to which they believed the stimulus belonged. To control for correlations between the participant's behavioural and neural responses the colour of the cue indicates the finger to use to indicate to which class they believe the stimulus belongs. A green cue indicates the index finger should be used for Class I and the middle finger should be used for Class II, while a red cue reverses this mapping.
The order in which trials were to be presented differed across runs and observers, and was generated such that despite the overlap in haemodynamic responses of adjacent trials there would be no statistical correlation between a trial and the one preceding.
3.3.4 fMRI Data Acquisition
The fMRI scanning sessions were performed at the Birmingham University Imaging Centre using a 3T Achieva Philips scanner with an eight-channel head coil. For localisation and visualisation of the functional data anatomical images were obtained for each participant using a sagittal three-dimensional T1-weighted sequence (voxel size=1 x 1 x 1 mm, slices- 175). Functional EPI images were acquired using a high-resolution gradient echo-pulse sequence (20 slices at 1.5 x 1.5 x 2mm resolution; TR: repetition time, 1500 ms; TE: time to echo, 34 ms; 4 dummy scans and 216 volumes acquired per run). The volume encompassed by these images was positioned to cover the occipital and posterior temporal cortex .
3.3.5 fMRI Data Preprocessing
Preprocessing was performed using Brain Voyager QX (Brain Innovation, Maastricht,
Netherlands). For the functional data this pre-processing included slice scan time correction, three-dimensional motion correction, linear trend removal, and temporal high-pass filtering. The resulting functional images were aligned and transformed into Talairach space. The first functional session was aligned to the anatomical data, and subsequent sessions scans were aligned to the first functional volume of the first session.
3.3.6 ROI Localisation
The regions of interest used for analysis were those identified in Kuai et al (2013) as involved in shape processing; these were the early regions V1 & V2, and the higher ventral regions V3v, hV4 and LO. Kuai et al (2013) identified the retinotopic visual regions used standard retinotopic mapping procedures as described by Sereno et al (1995) and Wandell et al (2007), while LO was defined as a sub region of LOC which showed stronger activation for intact images than scrambled images (Kourtzi & Kanwisher, 2001).
3.3.7 Multivariate Pattern Analysis
Eight multivariate pattern analyses were performed on the dataset described above. Four analyses were performed using the leave-c-out cross-validation procedure described in Section 3.2 both with no exclusion window and with exclusion window sizes of 6, 12 and 18 seconds. A further four analyses were performed using the same LCO cross-validation procedure and exclusion window sizes, however volumes were excluded using the random
exclusion method described in Section 3.2.4 in order to control for the effects of the reduction in trials available in the training set with the increase of exclusion window sizes. All leave-c- out crossvalidation analyses were performed with 100 folds and, barring the difference in exclusion methods, all eight of these analyses use the method described below.
Feature selection was performed with the threshold and ordering method described in Section 2.5.1. Voxels were first thresholded to include only those which showed significantly (p < 0.05, uncorrected) more activation during stimulus conditions than fixation. Next voxels were sorted in descending order of t-value with the first 150 voxels for each ROI being selected; in regions with fewer than 150 voxels available after thresholding all voxels would be included.
Normalisation was performed using the z-score method described in Sections 2.3 with the modifications described in Section 3.2.5. Sample estimation used the miniblocks method described in Section 2.4 with the modifications described in Section 3.2.1. The
haemodynamic delay was accounted for by shifting the fMRI time series by 3 volumes (4.5 s). Samples for each trial were generated by taking the mean of the values of the two volumes of the trial. These samples were then grouped into blocks of six trials of the same condition, with the a value for each block then being calculated as the mean of the six trials.
Classification was performed using a linear SVM model. The classifier was trained to categorise blocks as either being stimuli from Class I or Class II as indicated by the
participant in each trial. The potentially unequal number of trials from Class I and Class II in the training set was controlled for using the inbuilt cost-factor in SVMLight (Joachims, 1999)
which weights the cost of mis-categorising training samples.