Recognition of Brain Tissues, Coloring of Magnetic Resonance Images using SVM

The MRI (Magnetic Resonance Imaging) is based on a stimulation of hydrogen molecules in a tissue by a specific radio frequency

Abstract
The MRI (Magnetic Resonance Imaging) is based on a stimulation of hydrogen molecules in a tissue by a specific radio frequency, which when combined with a powerful space-variant time-invariant magnetic field, causes the molecules to emit EM radiation, which is detected and processed to create an image of the organ. Such images are usually representative of the quantity of hydrogen molecules in a specific pixel/ voxel, and the job of figuring out the type of tissue falls on the hands of the MR technician/physician. The system created has the ability of learning a classification of any given organ (specifically the brain, containing a large number of different tissues) based upon given MRI scans and a Phantom. It has a built in PCA (Principal Component Analysis) module which chooses the most informative scans out of a given database, which are fed into a multiclass SVM (Support Vector Machine), which creates a set of separation planes within the scan space, which determines the spatial location of each type of tissue. The system can learn to classify any type of tissue based upon MRI scans, Ultrasound, CAT scans etc.

The problem
Traditional MRI based diagnostics relayed on a set of different scans, each with a different scan parameters set, which emphasize different qualities of a given tissue. The analyzer must be highly trained in order to combine the date from at least 3 different contrasts, generally referred to as T1, T2 and PD. Therefore, the algorithm designed intended to create an SVM which will be fed these same contrasts, and based on a pre-determined phantom of the organ, will be able to create a set of rules, upon which future unknown scans could be analyzed.

12

 

34
Figure 1 – Three different contrasts of a brain scan and the corresponding Phantom

Different spatial-classifications were tested (such as Nearest-Neighbor, Center of Mass etc.), to initially determine the goals of the project (assuming Nearest-Neighbor should yield the best results), and eventually an SVM algorithm was written to classify the given scans.

The solution
5
Figure 2 – A Block Diagram of the System

The first stage of the system is learning the classification of the given tissues out of the given database, based on a predetermined phantom. The original phantoms used were obtained from BrainWeb – an online simulator created at Canada’s Mc’Gill University, contained 9 different tissues + background. The number of different types is variable, though the system is set to distinguish between 6 types. Next, the different types must be coded in a binary fashion, to allow a proper classification – The SVM is based upon a set of separation planes, and any type is defined in relation to these planes (+1 if it’s above, and -1 if it’s below). This codebook, which is created during the learning phase, is used during the operation of the system, and for the decoding part at the final stages. The next stage is finding the most significant contrasts from the given database using a Principal Component Analysis algorithm. This is done for both the training and testing databases (when and transformation parameters are determined during the learning phase and used later on, during operation).

6
Figure 3 – A 3D (3 contrasts) Spatial Representation of 6 Tissue Types after PCA

After the system had learned and created the classification rules, a new set of unknown scan is submitted, goes through a PC Breakdown according to the transformation parameters previously set, classification by the SVM, and decoding using the codebook. A postprocessing median filter may be used to smooth out local glitches, and may lower the error by up to 1%. A representative outcome is given below. Note that the error given relates to the entire picture. When calculation the error between Gray Matter and White Matter (the two types which are most important to distinguish between), the error is 6.5%. This is without median smoothing.

7
Figure 4 – Results for a given scan, without smoothing

Tools
All algorithms were written in Matlab.
Source code for the LS-SVM was obtained from: http://www.esat.kuleuven.ac.be/sista/lssvmlab/
Database of MRI scans was obtained from: http://www.bic.mni.mcgill.ca/brainweb/

Conclusions
The system implemented had succeeded with high performance in classifying brain tissues using the given database. As can be seen from the sample outputs above, less than 9% of the pixels were wrongly classified, and when keeping in mind the primary concern of differentiating between grey and white matter, even better results were obtained. These results are with very little difference from the “”Nearest Neighbor”” algorithm, which had been the scale according to which the system was tested. Further more, it has been noticed, that while the system was originally designed and tested on MRI scans, it maybe also be used to classify any other type of picture, given that a suitable phantom may be provided. For instance, it is with NO modification, that this exact system might be used to classify geological surveys based Ultrasound scans, and classify different soils. Further uses may be suggested, all based on the demand to classify pictures to given categories according to a given database of phantoms and scans.

Acknowledgment
We are grateful to our project supervisor Michael Bronstein for his help and guidance throughout this work. We are also grateful for all the help and support provided for us by the VISL staff and personal and the Ollendorff Minerva Center that supported this work.