5th World Congress on Industrial Process Tomography, Bergen, Norway New Hull-Voxel Approach to Image Reconstruction from Limited Projections and Views V. Vengrinovich 1 , S. Zolotarev 1 , W. Schlegel 2 , B-M. Hesse 2 and J. Hesser 3 1 Institute of Applied Physics, Minsk, Belarus, Email: veng@iaph.bas-net.by 2 German Center for Cancer Research(DKFZ), Heidelberg, Germany 3 Mannheim University, Mannheim, Germany ABSTRACT We discuss the problem of image reconstruction from limited views and sparse projections in order to achieve a more accurate diagnosis of the patient’s morphology. As a priori assumption, we demand that the object consists of regions with constant attenuation coefficient. Hereby, the number of such regions is not restricted. The Hull-voxel reconstruction concept is a new method based on this a priori information and thereupon makes up for the lack of projections data. In the first reconstruction phase the Bayesian approach is applied, while at the second phase the so called hull-voxel iteration reconstruction algorithm is developed, which provides the iterative fitting of regions’ surfaces by global optimization technique. The results are evaluated based on simulated Shepp-Logan and experimental medical phantoms respectively. Keywords Tomography, limited data, planar geometry 1 INTRODUCTION Image reconstruction from incomplete and noisy X-Ray projections yields increasing attention, intensified by clear practical demands for reduced radiation dose; real time tracking of e.g. a tumor, organ or a machine part. This trend towards the development of image reconstruction algorithms and tools which would operate with incomplete data sets is encouraged by the transition from fan to cone beam data acquisition and local tomography cases. If a desired Region of Interest (ROI) is entirely “seen” from sufficient set of x-ray projections (circular trajectory and filled Radon space), then one of the available techniques, either based on the classical Radon transformation (see for fundamentals e.g. Natterer, 1986), or iterative approaches (see the tutorial Press et al. 1992), can be used for reconstruction. However, in a case of limited observation angles or limited number of projections, or both, the problem becomes extremely ill-posed. Traditional algorithms are unable to achieve an artefact-free reconstruction and often lead to severe image blurring in the direction of X-ray propagation. Blurring increases with fewer projections and smaller viewing angles. The limited angle problem is much more critical than sparse projections. The Bayesian technique, complemented by efficient prior knowledge, is recognized as one of the most powerful tools for image reconstruction from limited data sets (Fessler, 2000). Statistical properties as prior for Bayesian reconstruction, whilst being very productive for binary images (Hanson, 1987; Vengrinovich et al. 1999a; Vengrinovich et al. 1999b), are less efficient in compensating for missed data, especially for the case of low contrast medical images. For certainty, we consider the case of a planar tomographic geometry with a patient located in a space between parallel planes, each (or at least one) being a circular trajectory, one for an x-ray source and the other for a planar detector respectively (Figure 1). The typical application is a portal-type data acquisition setup, with the position of kV source at the opposite side of an object with regard to MEV accelerator target (Hesse, 2003) (Figure 1). The acquisition geometry (Figure 1) is provided by a circular (around vertical axes z) trajectory of an x-ray source. In the case of a circular trajectory, the admissible solid angle for object observation cannot exceed 90 0 and thus one has to deal with a limited angle problem. In this article we present a dual-phase constrained 3D image reconstruction technique of a region of interest (ROI), which can overcome an image degradation resulting from limited data. This technique is 208