A Simple Application Framework for Interactive Segmentation Systems S.D. Olabarriaga D. Koelma A.W.M. Smeulders Intelligent Sensory Information Systems, WINS, University of Amsterdam, Kruislaan 403, 1098 SJ Amsterdam, The Netherlands silvia koelma smeulders @wins.uva.nl Keywords: Interactive Segmentation, Image Segmentation, Software Architecture, Medical Imaging Abstract This paper presents a simple software architecture for interactive segmentation (IS) systems. We analyze system requirements based on a generic model for IS processes and a typical application. The proposed sys- tem architecture is presented with focus on module or- ganization and the interface for communication among them. Features not present in programming environ- ments for development of image processing applica- tions are detailed: the IS general process model, the image visualization window, and the graphical input tools. 1 Introduction Interactive Segmentation (IS) has been recently proposed in [1] as a possible approach to the difficult problem of medical image segmentation. User inter- vention is suggested as an extra source of information for cases where automatic techniques do not achieve acceptable results. Two main cases for failure of au- tomatic procedures were identified in [1]: (a) the im- age does not contain enough visual evidence and (b) the object at hand deviates from the expected shape model. In the approach described in [1] user interven- tion consists of corrections to the segmentation result using specialized IS tools which also inform the rea- son for failure. These tools do not directly update the output of segmentation, but instead use the provided information to drive the automatic segmentation pro- cess to a satisfactory solution. The resulting segmen- tation is thus a compromise between the automatic pro- cedure, based on the image data and the shape model, and the user input expressed as corrections with associ- ated semantic. Results obtained with IS techniques are expected to display more objectivity and reproducibil- ity than automatic processing combined with manual correction. In [2] we propose a general model for IS which of- fers not only a promise for more objective and repro- ducible results, but also allows the design of an ap- plication framework for easy prototyping of new IS techniques. In this paper we first present the general IS model with an example of application. Next we analyze the requirements for a comfortable environ- ment to develop such applications. We then describe the proposed system architecture and the basic services with emphasis on the issues not yet covered in the cur- rently available software infrastructure. 2 General IS Model For the moment we concentrate on boundary-based segmentation techniques, particularly on deformable contours [3]. In these techniques an initial contour is deformed by image forces while keeping shape con- straints. We formulate IS as an iterative process in which a curve (the object’s contour) is deformed with user participation until it displays a desired quality condition . The quality measure is com- puted as a sum of terms based on: image data: fitness of the contour to the visual evidence of edges present in the image data (e.g. gradient of image intensity); shape model: similarity between the contour and some a priori shape model measured at global and local scales (e.g. match to template and curvature at points in the contour); user input: interpreted as corrections to the cur- rent segmentation result which affect the compu- tation of (e.g. the relative importance of the shape- and image-derived terms).