so i thought i could fill you up on what i did on my thesis.
i expect you all biology freaks know about MRI. now given an MR image, you identify the different parts of the brain. major tissues are gray matter (GM), white matter (WM) and Cerebro-Spinal Fluid (CSF). MR images are not like xrays where you take a glance and know which part is which. the tissues are very soft that they sort of mix together and its hard to know where the boundaries are. so segmenting the image into these tissues is a problem. its refered to as MRI segmentation.
typically, three scans of the same object are made. the have different characterisitics and they give somewhat different images of the same object. the importance is that one part of the object may not be easily seen in one image but another image may show it clearly. the first row in the image below shows an MRI scan. these are three scans of the same “slice” of brain. the second row shows the three tissues we talked about. its not like one image corresponds to the image just above it. all three images (a)-(c) are used to identify the tissues (d)-(f).
now in my field there are methods which are used to group related things together. in this case, “related things” would be the tissues. so given a pixel (a very small patch of the image), the task is to say which tissue it belongs to. there already cool ways of doing it using computers but i was trying to make an improvement over those. i’ll save you the details
rows three to five show the results from experiments. by the way, the second row is a “true” segmentation which is being used to evaluate how good the results are. the third row shows results from my initial model. which kind of sucks. so i tried to improve it. i started with some other model which gives results on the fourth row. i improved on that and got the fifth row. which is good (compare with the second row)
not to say i have made a breakthrough
there are excellent methods for doing this which give results much better than this. plus, this image is simulated image and it has some nice properties which make it easy to deal with. the significance of my approach is that some improvements can be made by doing the stuff i did. and i’m concluding that these things should be done on the arif models currently being used.
gebito?






