Optimal feature extraction for the segmentation of medical images

Many image segmentation algorithms use a small local area around each pixel for the extraction of features, in order to minimise the effect of image anomalies. The main drawback of this approach is its generation of classification errors at region boundaries, where the local area can contain pixels from more than one region. In this paper, a novel method of determining the optimal position of the local area for feature extraction is presented. The proposed technique avoids overlap into adjacent regions by examining the intensity gradients of neighbouring pixels and shifting the area for feature extraction accordingly. The improvement obtained using this technique is demonstrated on a variety of MRI medical images.