Application of Information Theory to Improve Computer-Aided Diagnosis Systems
Abstract
Mammographic Computer-aided diagnosis (CAD) systems are an approach for low-cost double reading. Though results to date have been promising, 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 common elements of most CAD systems. This research project looks to apply principles from information theory to build statistical models for CAD systems. Specifically, we develop a framework for building hierarchical pattern recognizers based on the minimum description length principle (MDL). Under the first year of this project we have developed a framework for building generative hierarchical image probability (HIP) models. Since the HIP framework is a generative model which directly models the probability of the image given the image class, it is well-suited to compression and application of MDL. We have started building HIP architectures using the MDL selection criteria. For example we have used predictive MDL (pMDL) to select the number of segmentation labels at each level of the pyramid. Finally, under our first year's effort, we have also expanded our hierarchical modeling framework to include both microcalcification and mass detection.
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 01, 1999
- Accession Number
- ADA383361
Entities
People
- Paul Sajda
Organizations
- Sarnoff Corporation