A multi-resolution segmentation scheme and its application to edge detection in cardiac MR image sequences

We present a multi-resolution segmentation scheme for magnetic resonance images based on vector quantization of a block partitioned image followed by a relaxation labeling procedure. By first searching a coarse segmentation, the algorithm yields very fast and robust performance on images that are inherently noisy, and is particularly suitable for processing of a sequence of images. The algorithm defines feature vectors by the local histogram on a block partitioned image, and approximates the local histograms by normal distributions. This is a suitable feature extraction for medical images since most are tone images with short-term correlation. Within this framework, the least relative entropy is chosen as the meaningful distance measure between the feature vectors and the templates. The segmentation is performed by a block-wise classification-expectation algorithm, and is improved by a multi-resolution procedure. The scheme is applied to cardiac MR image sequences and results are presented to show that the algorithm successfully extracts the endocardial contours and that temporal processing significantly improves the edge detection performance and can avoid local minima problems.