Constantino Carlos Reyes-Aldasoro and Abhir Bhalerao, Classification of Human Knee Data from Magnetic Resonance Images (April 10, 2002).
This report considers the general problem of segmentation of Magnetic Resonance Images. The final objective is to correctly assign a unique label or class which represents an anatomical structure to every pixel or voxel in a data set. The images analysed describe a human knee scanned by Magnetic Resonance (MR). A brief description of the anatomy of the knee and physics of MR imaging is given. A review of image segmentation approaches, focusing on multiresolution and texture segmentation, follows. The first segmentation technique implemented here is grey level thresholding, which is later improved by adding two other descriptors of the images: standard deviation and a moment from the co-occurrence matrix. Frequency analysis through sub-band filtering is proposed as a way to improve the description of the textural regions and boundaries between anatomical regions. Comparative results of the different techniques are presented and finally conclusions and future work is proposed.