From statistical pattern recognition to decision support using
probabilistic networks -- similarities and differences

Statistical pattern recognition and the framework of probabilistic networks (belief networks) both build upon Bayesian statistics. While techniques from statistical pattern recognition - classifiers, neural networks and support-vector machines - are highly applicable to classification problems and low-level image processing tasks based on continuous features, the framework of probabilistic networks has been developed specifically for high-level probabilistic reasoning with discrete variables. Pattern recognition techniques and probabilistic networks in a sense are complementary tools that are linked by their common basis offered by Bayesian statistics.

A review of applications of neural networks and related pattern-recognition techniques in image processing has indicated that these techniques have been applied successfully to low-level pattern recognition tasks such as object recognition and texture segmentation. However, the incorporation of domain knowledge in image processing applications is very limited (partly due to the tradition within the field), and mostly boils down to selecting a useful set of features and an appropriate prior (class) distribution. In medical applications, however, domain knowledge can often aid the correct interpretation of an image. Moreover, such knowledge is required for, for example, therapeutic decisions which are based upon the investigation. In this context, the result of a low-level image processing algorithm is one of several factors that come into play. Variables like age, sex or risk factors such as smoking co-determine how the image should be interpreted and which medical decisions that are the best to make. The framework of probabilistic networks makes it possible to include such clinical variables in a stochastic model which can aid the interpretation of images and the subsequent decisions. The framework offers a mathematical formalism and a set of algorithms that make it possible to model highly complex decision problems and compute the probabilities of all outcomes of interest.