Brain Functional Distinguishability Measure

Some more thoughts about this project. (hand writing pdf)

Motivation and definition (pdf)

The specificity toolbox for MATLAB

The specificity toolbox for MATLAB is made available here.

How to use the toolbox:

1. Download the toolbox here, and extract it.
2. In the extracted directory, there will be some demo files.
3. You might want to use the example data file, which can be download from here.
4. Make sure you point to the data file correctly.
5. Run the demo files.

How to use the toolbox for fMRI research:

1. Download the MVPA demo file, and make sure you include the path to the toolbox and the data correctly.
2. Run the demo file and get some results and plots.

Experiment1: Preliminary results on separability metric on ventral temporal (VT) cortex

I use "averaged distinguishability" V as a metric to figure out which voxels are "versatile" or "non-informative".

versatile: A voxel is said to be "versatile" when its beta response can distinguish any pair of class.

non-informative: A voxel is said to be "non-informative" when its beta response cannot distinguish any pair of class.

In this experiment, I calculate the distinguishability V_i for all voxel i in the VT cortex (whose mask is given by the Haxby's data set) and the whole brain. A histogram of V is shown below in the figure.

There are some facts worth talking about:

1. Most of voxels have medium score (V<0.3)
2. Very few voxels are versatile (V>0.8) even in VT cortex
3. There are more voxels having medium score than having very low score (V<0.1) in VT cortex, suggesting that voxels in VT cortex are quite informative

The fact that most of the voxels in VT cortex having pretty low V suggests that a certain voxel can specifically distinguish some pairs of classes only. Adding the previous results stating that we can achieve about 86% accuracy when using all the voxels, we can infer that voxels work jointly to classify class label. In other words, a single voxel is not a good feature to classify task, but when they are used jointly with others, they can be very good features. This finding is consistent with Haxby's stating that the information is distributed.

Experiment2: Preliminary results on separability metric on whole brain

Now we conduct the same experiment as the one before but with the whole brain. It is clear from the histogram that majority of the voxels are non-informative and only a few voxels are informative. From the results, most of the informative voxels locate very close to the one in VT cortex. However, there are some other regions in the brain that are pretty informative rather than VT cortex.

Experiment2a: Correlation between V and accuracy and MI

It is nice to see how much the measure V is correlated to other measures, for instance:

the independent single voxel accuracy and the MI for each voxel url

Experiment3: Class-specific measure

In this experiment, I came up with a measure to compute the distinguishability of a specific class. link

Experiment4: Classification result using only non-informative voxels in VTC and whole-brain

In this experiment, I came up with a measure to compute the distinguishability of a specific class. link

This is a very interesting experiment as it will be used to proof the information in the brain is distributive link.

Experiment5 Classification result using only versatile voxels (V>0.5) in VTC and whole-brain

In this experiment, I came up with a measure to compute the distinguishability of a specific class. link

Experiment6 Compare the classification accuracy between k-nn and k-random-neighbor (k-rn) on all voxels in VTC and whole-brain

In this experiment, the results are shown here

Next

1. We will use this distinguishability matrix as a feature to do unsupervised clustering. I want to try hierarchical clustering, spectral clustering.
2. What would be the classification result when using the high-V voxels combined together. The results should be compared with the case when MI is used.
3. Another interesting idea is how we can find a system/network of voxels such that each individual voxel has a low V value but gives a high classification results when combined.
   1. We might start from first asking what are the single voxels that yield the best accuracy?
   2. We then ask what are the two-voxel pair to give best accuracy?
   3. What are the three-voxel triple to give the best accuracy?
4. Finally, we should connect this idea to social network. A group of voxels can be called "community". We can do the clustering based on their link rather than using the nodes. Refer to Yong-yeol's Nature paper.