Block-wise segmentation via vector quantization for medical image analysis

We present a new segmentation and modeling scheme for medical images based on vector quantization which yields very fast responses and can avoid local minima in the computation. The feature vectors are defined by the local histogram on a block partioned image, and the local histograms are approximated 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 then performed by a block-wise classification-expectation algorithm, and is improved by a multiresolution procedure. The performance of this learning technique is tested with both simulated and real medical images, and is shown to be a highly efficient segmentation scheme