A Bayesian Approach to in vivo Kidney Ultrasound Contour Detection Using Markov Random Fields Marcos Mart´ ın and Carlos Alberola ⋆ ETSI Telecomunicaci´on. Universidad de Valladolid Cra. Cementerio s/n, 47011 Valladolid, Spain {marcma,caralb}@tel.uva.es Abstract. Automatic detection of structures in medical images is of great importance for the implementation of tools that can obtain accu- rate measurements for an eventual diagnosis. In this paper, a new method for the creation of such tools is presented. We focus on in vivo kidney ultrasound, a target in which classical methods fail due to the inherent difficulty of such an imaging modality and organ. The proposed method operates on every slice by detecting kidney contours under a probabilis- tic Bayesian framework. We make use of Markov Random Fields ideas to model the problem and find the solution. A computer easy-to-use interface to the model is also presented. 1 Introduction Neonatal hydronefrosis is a disease of great relevance in the fetus and newborn children. It consists of an enlargement of the renal pelvis and calyces. Its early diagnosis is a common task thanks to the use of echography, both during the pregnancy or in the newborn and is becoming the more frequent prenatal urologic diagnosis. Echographical analyses permit determining whether this or other urological diseases are present; the current inspection process is as follows: after scanning an adequate slice, the specialist manually adjusts —helped with cursors— an ellipse to the guessed external boundary of the kidney. The system approximates the kidney volume as the volume of the ellipsoid generated by rotating the sketched ellipse about its main axis. The pelvis volume is determined similarly. From those approximations, the specialist reaches a diagnosis using tabulated tendency data. An automatic or semiautomatic segmentation tool will be, in our opinion, of valuable importance, not only for the determination of the kidney and pelvis contours by means of a more accurate method than the one described above, but also for automatically propagating those contours to the rest of the slices of the volume scanned. With such an approach, the volume estimates are expected ⋆ The authors acknowledge the Comisi´on Interministerial de Ciencia y Tecnolog´ ıa for research grants 1FD97-0881 and TIC2001-3808-C02-02, and Junta de Castilla y Le´ on for research grant VA91/01. T. Dohi and R. Kikinis (Eds.): MICCAI 2002, LNCS 2489, pp. 397–404, 2002. c  Springer-Verlag Berlin Heidelberg 2002