Reliability analysis framework for computer-aided diagnosis systems in mammography
Piotr A. Habas · 2007
The main focus of this dissertation is the formulation, implementation and evaluation of a reliability analysis framework for computer-aided diagnosis (CAD) systems in mammography. The presented technique enhances CAD systems by providing them with the ability to assess the reliability of each individual decision they make. The query-specific reliability assessment is achieved by measuring the local accuracy of a CAD system with known cases highly similar to the query. It is expected that CAD decisions assigned higher reliability scores are more accurate. The proposed reliability analysis technique was thoroughly tested using three decision models of different nature and structure and a mammography dataset of 1337 regions of interests. Statistical analysis of experimental results revealed that the proposed query-specific reliability measure is a strong predictor of the CAD performance for the query case. As this observation was consistent across all decision models investigated in this study, it is concluded that the presented reliability analysis framework is independent of the nature and the structure of the underlying decision model and can be applied with any type of classifiers. This dissertation also addresses the issue of integration of reliability information into the medical decision making process. The proposed approach builds upon the reliability analysis framework and advances it to a clinically meaningful risk stratification strategy. Finally, this dissertation introduces a probabilistic adaptation of the proposed reliability analysis framework for increasingly popular case-based reasoning systems.