A knowledge based segmentation method applied to medical radiographs

The problem of identifying an individual´s skeletal age from a radiographic image of the hand and wrist is of importance to radiologists. From a radiograph a person´s growth and development can be assessed and examined for abnormal conditions. At present, a person´s age is estimated either by visual comparison of radiographs using an atlas compiled by Greulich and Pyle (1959) or by combining visual and textual descriptions as detailed by Tanner and Whitehouse (1975). Both methods are time consuming and results are often subjective, see Cole (1988). Computer analysis of skeletal age would overcome the above restrictions. Three separate levels of processing may be identified to achieve this goal; bone boundary detection, bone identification and classification. The subjects of the paper are (a) detection of bone boundaries using closed contour edge detectors to segment the image and (b) clustering to form regions representing the bone detail. Segmentation methods which rely on first derivative edge detection have shown to be ineffective due to the occurrence of gaps in the resulting edge maps. One approach is to prevent such gaps from occurring by improving the method of detecting edges. Three solutions to the problem of closed contour edge detection are proposed: an anisotropic Laplacian of a Gaussian operator; a space limited Laplacian of a cosine operator; and a gradient tracking scheme