CIF: Small: Theory of Multiresolution Classification with Bases and Frames

Carnegie Mellon University, Pittsburgh PA

Investigators

Abstract

Recent advances in imaging of biological systems at all scales, from molecular and cellular up to organ levels, have given biologists and clinicians opportunities to observe processes and interactions at a never-before-seen level, leading to the collection of huge amounts of high-dimensional data. As a result, the visual inspection of these data sets,already error-prone, nonreproducible and subjective, has become impractical as well. There is thus an acute need for the development of systems to both automate this analysis, as well as mine interactions not visible to the human eye. The task of classification has been at the heart of several of the group's projects in the past few years, including the determination of developmental stages in fly embryos,the recognition of H&E-stained tissue types in stem-cell teratomas, and the diagnosis of otitis media. As an accurate and efficient algorithm for automated classification would have been of great use to biologists and clinicians, a multiresolution (MR) classification algorithm was developed and, in each of the problems, consistent trends emerged: 1. MR classification always performed better than the no-MR version; 2. Redundant MR transforms frames, always performed better than the nonredundant ones bases. This consistency across data sets and applications indicates that MR has the power to make a significant impact on biomedical image classification performance. The investigators thus study MR classification to gain fundamental understanding of its underpinnings, in particular, the following two questions: 1. When/why does the MR classification work? 2. When/why does the MR frame classification work? These questions are approached by setting up a measure-theoretic theory of classification as a mathematically rigorous framework within which to pose and investigate real-world classification problems.