ABSTRACT
A medical image dataset is the starting point for important epidemiological and statistical studies. In fact, it can be used to develop and test algorithms for computer-aided detection (CAD) systems, as the development of a CAD system is strictly related to the collection of a large dataset of selected images, for teaching and training medical students or as an archive of rare cases (Tangaro et al. 2008) (see Section IV, Chapter 59). This is true also for mammography, in particular after the worldwide spread of mammographic screening programs (see Section IV, Chapter 60). When the evaluation of a CAD algorithm begins with a retrospective evaluation of cancer cases (Heath et al. 2001), such preliminary evaluation is more time and cost effective that a prospective evaluation in a clinical setting. The task of obtaining the data for a retrospective CAD performance evaluation at a mammography center may be time consuming and expensive to achieve. When investigators of CAD methods utilize the resources in a common database, rather than using their own data, this expense may be decreased, and much more may be learned by a performance evaluation. In fact, provided the same data, the same performance measure and the same train and test methodologies are followed, results from different algorithms can be compared to find the relative strengths of each algorithm. This led to the development of new or combined approaches to the problem that yield superior performance (Brake and Karssemeijer 1998).
