Automatic segmentation of clinical structures for RTP: Evaluation of a morphological approach

This paper describes a morphological segmentation technique developed to assist clinicians and radiologists in conformal RTP. The system uses a watershed transformation to give an initial over-segmentation of a 2D CT; this is then refined by homotopy modification with markers identified automatically from the image histogram and the final segmentation is represented by a region adjacency graph which may be interactively edited by the user. To evaluate the system, 77 CT images from 11 patients have been segmented and the results compared against those produced by 5 clinical experts. The approach has enabled the quality of the automatic segmentation to be assessed in terms of the overall expert intra/inter variability. It has been proved good agreement between computer generated and expert manual delineations.