Unsupervised fMRI Analysis

Recently machine learning methodology has been used increasing to analyze the relationship between stimulus categories and fMRI responses [2, 14, 15, 11, 13, 8, 9, 1, 12, 7]. Here, we introduce a new unsupervised machine learning approach to fMRI analysis approach, in which the simple categorical description of stimulus type (e.g. type of task) is replaced by a more informative vector of stimulus features. We compared this new approach with a standard Support Vector Machine (SVM) analysis of fMRI data using a categorical description of stimulus type. The following study differs from conventional unsupervised approaches in that we make use of the stimulus characteristics. We use kernel Canonical Correlation Analysis (KCCA) to learn the correlation between the fMRI volume and the corresponding stimulus features presented at a particular time point. CCA can be seen as the problem of finding basis vectors for two sets of variables such that the correlation of the projections of the variables onto these basis vectors are mutually maximised. KCCA first projects the data into a higher dimensional feature space before performing CCA in the new feature space. To classify a new stimulus we compute the kernel between the test and the training samples and project that kernel onto the learned semantic space. The output gives a score which can be thresholded to allocate a category to each test example. CCA [3, 4] and KCCA [5] have been used in previous work for fMRI analysis using only conventional categorical stimulus descriptions and did not explore the possibility of using the complex characteristics of the stimuli as the base for the feature selection in the fMRI data. The fMRI data used is from an experiment in which we studied the responses to stimuli designed to evoke different types of emotional responses (pleasant/unpleasant). The pleasant images consisted of women in swimsuits while the unpleasant images comprised image of skin diseases. For experiment details see [7]. Each stimulus image was represented using Scale Invariant Feature Transformation (SIFT) [10] features, which are used in a similar fashion to the bag-of-words (word-frequency) models for text documents. The complete dataset consisted of 1344 fMRI volumes and their corresponding imagestimuli. From this dataset, we used 670 fMRI volumes (517, 845 voxels) and their corresponding image-stimuli for training and the remaining 674 for testing. In both training and testing datasets half the stimuli were Pleasant and half Unpleasant, hence a random classifier would achieve 50% accuracy. The experiments are repeated ten times, each with a random permutation of the training-testing set. We use linear, centralised, kernels in our KCCA framework. The KCCA regularisation parameter [6] is computed using 2−fold cross validation on the training data. KCCA achieves an accuracy of 81.34%± 3.29% while the supervised SVM attains an accuracy of 90.28% ± 0.93%. These classifications accuracy are significantly different t-test p << 0.0001. Thought in both cases the results are much higher than the chance expectation. The unthresholded weights vectors for both methods are displayed in 1. The similarities in the patterns of the weight vectors obtained by the two classification methods provides reassurance of the validity of the KCCA analysis, which is effectively obtaining the categorical stimulus type directly from the data. SVM: Non-threshold Weight Vector (UnpxPle) Using single volumes as examples Accuracy ple 73.8 % Accuracy unp 73.0 % I also got an accuracy of ~90% when I averaged the examples within the blocks to increase the SNR (I have different SVM results I don’t know what you want to show for comparison)