This paper is on automated visual inspection of tablets that may, in contrast to manual tablet sorting, provide objective and reproducible tablet quality assurance. Visual inspection of the ever-increasing numbers of produced imprinted tablets, regulatory enforced for unambiguous identification of active ingredients and dosage strength of each tablet, is especially demanding. The problem becomes more tractable by incorporating some a priori knowledge of the imprint shape and/or appearance. For this purpose, we consider two alternative automated tablet defect detection methods. The geometrical method, incorporating geometrical a priori knowledge of the imprint shape, enables specific inspection of the imprinted and non-imprinted tablet surface, while the statistical method exploits statistical a priori knowledge of tablet surface appearance, derived from a training image database. The two methods were evaluated on a large tablet image database, consisting of 3445 images of four types of imprinted tablets, with and without typical production defects. A 'gold standard' for testing the performances of the two inspection methods was established by manually classifying the tablets into good and five defective classes. The results, obtained by ROC (receiver operating characteristics) analysis, indicate that the statistical method yields better defect detection sensitivity and specificity than the geometrical method. Both presented image analysis methods are quite general and promising tools for automated visual inspection of imprinted pharmaceutical tablets.

No abstract

No abstract

This paper presents a framework for the segmentation of anatomical structures in medical imagery by connected statistical models. The framework is based on three types of models: first, generic models which operate directly on image intensities, second, connecting models that impose restrictions on the spatial relationship of generic models, and third, a supervising model that represents an arbitrary number of generic and connecting models. In this paper, the statistical model of appearance is used as the generic model, whiles the statistical model of topology, obtained by applying principal component analysis (PCA) on aligned pose and shape parameters of the generic model, is used as the connecting model. The performance of such connected statistical model is demonstrated on anterior-posterior (AP) X-ray images of the hips and pelvis and compared to the modelling by one and six unconnected generic models. The most accurate and robust results were obtained by two-level hierarchical modelling, wherein connected statistical models were used first, followed by unconnected statistical models.