An optimized statistical approach to magnetic resonance image segmentation
The usage of image segmentation techniques for the purpose of analysis of medical imagery is an essential facet in medical diagnosis. These techniques provide a potential Computer-Assisted Diagnostic (CAD) tool for today's medical industry. The aim of this research project is to use the DA algorithm for segmenting three main tissue classes from the MRI brain images in order to enable the human interpreter to recognize any abnormalities (such as lesions,) which have been enhanced by this procedure. DA is based on the minimization of a cost function that includes Shannon's Entropy in addition to the expected distortion. The mass-constrained version of the algorithm, as implemented here, successfully eliminates code-vector repetition, thus enabling faster convergence of the cost.
The results indicate that lesion isolation in the case of simulated images requires further steps of pre- and post-processing of the source imagery. These methods successfully isolate the lesion areas in a specific case with confirmed MS in an advanced stage and show abnormalities in brain tissues for a suspected case. The analysis of the segmentation results indicates potential usage of DA in the area of medical image segmentation and also points towards future research and further improvement. In conclusion, the feasibility of this methodology to recognize MS lesions has been tested in this thesis, using both simulated and clinical MRI data.