Efficient segmentation in MRI applying discrete wavelet transform and topology preserving neural networks

The paper investigates a novel feature extraction approach to MRI segmentation based on identifying the critical image edges by formulating the problem as a two-stage unsupervised classification task using a modified Kohonen's self organizing feature map (SOFM) along with independent component analysis (ICA). The main goal of such a research effort is to better identify abrupt image changes without increasing the presence of noise in the resulting image. The suggested methodology is based on novel discrete wavelet descriptors involving the discrete k-level 2-D wavelet transform applied to sliding windows raster scanning the original image. The proposed two-stage classification scheme applied to such wavelet descriptors and using a modified vector quantizing self-organizing feature map (SOFM) and ICA analysis is compared with a corresponding two-stage scheme involving SVD analysis and the widely used SOFM, trained with Kohonen's algorithm. The feasibility of this novel two-stage proposed approach is studied by applying it to the edge structure segmentation problem of brain slice MRI images. The promising results presented in the experimental study illustrate a performance favorably compared, also, to that of traditional Sobel edge detectors supported by usual contour tracing methods.