Application of Information Theory to Improve Computer-Aided Diagnosis Systems

摘要

Computer-aided diagnosis (CAD) systems for mammography are an approach for low-cost double-reading. Current systems often suffer from unacceptably high false positive rates. Improved methods are needed for optimally setting the system parameters, particularly in the case of statistical models and neural networks which are a common element of most CAD systems. This research project looks to apply principles from information theory to build statistical models for CAD systems. Under the second year of this project we have further evaluated our generative hierarchical image probability (HIP) model, trained using information theoretic model selection. Our results show that HIP can reduce false positive rates by 30% for a data set constructed using The University of Chicago CAD mass detection system. We have also demonstrated the generative utility of our HIP model. We have synthesized regions of interest (ROIs) using the HIP model, enabling one to gain an intuition into the structure the HIP model learns for representing the two classes. Finally we have used the generative structure of the HIP model to detect novel examples-examples that significantly differ from the training data. We have shown how novelty detection can be used to generate confidence measures for improved ROC performance.