In this article we discuss an article by ( Kuncheva et al) [1] named “Random Subspace Ensembles for fMRI Classification” in which Artificial Intelligence techniques such as SVM, Random Forests are combined with probabilistic modelling to predict efficiently the brain functioning based on MRI scans. This research ( Kuncheva et al) [1] talks about application of Artificial Intelligence for functional MRI (magnetic resonance imaging) based classification of certain images and its influence to human brain. The following data was analyzed. This data can help one understand much more of the aims and indeed do it quite clearly.
- Haxby Data [Provided by MATLAB [2] toolkit too]. The categories of images shown to candidate human were eight in number and included human face, certain animals, objects made of metal, brick and plastic such as chair, scissors etc. to certain absurd images. The person who nominated for experiments was noted on brain activation after every 2.5 seconds while the images were displayed on screen for 22.5 seconds with a gap in each image being 12.5 seconds. Along with this data, image type or category, to be more specific is also stored.
- Bangor 1 Data: This data was collected from a 35 year old man on 3 Tesla Philips Achieva MR scanner. The categories of image shown to the man were positive images and negative images which were randomized. Each image was shown for 6 seconds and label of image (2 classes here) were saved along with data. Emotions were noted on scale of 9 varying from worse to best.
Similarly Bangor 2 Data etc. was collected by the participant under different goals and different inputs.
Random Subspace (RS) ensemble method for functional MRI classification works on the criterion that the number of features in these kind of problems is too high and number of examples are comparatively very low. The authors proposed a ratio to determine how many features per object of data is pertinent. The high dimensionality of problem makes it evident to use techniques that can work in tandem with huge datasets.
The technique works on selecting random features from data, the selection is made on selection without replacements and have been mathematically proved that when the number of such samples is fixed number L, can be considered to as being selected with replacements.
Now, the idea of this research is that- L number of samples of features are selected. In each of these L choices of features, the aim is to search key features, or important features. The sample that has more number of important features is considered important. Once a suitable number of samples have been collected, note features are subsampled here, then classification is performed, using classifier such as SVM, Random forest. For each of the sample, the class is determined for a testing sample and the final class of sample is determined by majority votes. (Majority Votes: Choose the category of sample as the one that has maximum count for the given testing sample).
References
- Ludmila I. Kuncheva*, Juan J. Rodríguez, Catrin O. Plumpton, David E. J. Linden, and Stephen J. Johnston, Random Subspace Ensembles for fMRI Classification, IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 29, NO. 2, FEBRUARY 2010.
- MATLAB: https://www.mathworks.com/products/matlab.html
You must log in to post a comment.