A New Hybrid Algorithm Based on Watershed Method, Confidence Connected Thresholding and Region Merging as Preprocessing for Statistical Classification of General Medical Images

Segmentation of morphology in medical image data is a highly context specific and differs from various imaging modalities, necessitating the use of sophisticated mathematical models and algorithms to achieve good results. In this work an algorithm is presented for presegmentation of general medical input data, based on a watershed-segmentation strategy utilizing both, original intensities and derived gradient magnitudes for region growing. The number of resulting pre-classified regions is iteratively reduced to a user-defined threshold using merge metrics, accounting for the similarity of intensity profiles of two neighboring regions to merge, as well as the height of the gradient barriers to overcome and geometric aspects like sphericity and size of the border area with respect to the total region size. Based on such a context-independent pre-segmentation, the resulting manageable number of regions can be further merged and classified, utilizing texture features and a priori statistical models. Results are presented from brainweb database.