Determining position and fine shape detail in radiological anatomy

Abstract

In this paper a method is proposed that identifies bone positions and fine structure of bone contours in radiographs by combining active shape models (ASM) and active contours (snakes) resulting in high accuracy and stability. After a coarse estimate of the bone position has been determined by neural nets, an approximation of the contour is obtained by an active shape model. The accuracy of the landmarks and the contour in between is enhanced by applying an iterative active contour algorithm to a set of gray value profiles extracted orthogonally to the interpolation obtained by the ASM. The neural nets obtain knowledge about visual appearance as well as anatomical configuration during a training phase. The active shape model is trained with a set of training shapes, whereas the snake detects the contour with fewer constraints and decreases the influence of a priori knowledge in a controlled manner. This is of particular importance for the assessment of pathological changes of bones like erosive destructions caused by rheumatoid arthritis. © Springer-Verlag Berlin Heidelberg 2003.