Data Mining for Large Data Sets of Shapes Deformations

University Of Houston, Houston TX

Investigators

Abstract

This project aims to reconstruct and quantify the dynamic deformations of soft organ shapes which are routinely observed by clinicians or researchers in medical imaging data. Our mathematical and computational approaches can potentially impact many domains, ranging from computer aided medical diagnosis, to automatic shape recognition in computer vision, as well as automated clustering, classification, and fast retrieval of biomedical image sequences. We will study movies recording dynamics of three dimensional biomedical shapes, for instance in live echo-cardiographic imaging of patients beating hearts. We will quantify shape distortions by computing "elastic distances" between deformable shapes and by large numbers of "strain values" evaluating local deformation of tissues. In turn these deformations characteristics enable the use of artificial neural networks for automatic classification and clustering of deformable shapes. Our work is motivated by the increasing availability of large databases of movies recording in 3D the dynamic deformations of "soft" shapes, such as human organs. We aim to generate quantified comparison between any pairs of movies recording the dynamic deformations of similar biomedical shapes, such as the mitral valves of cardiology patients. We will use one large seed set of actual echo-cardiographies of mitral valves dynamics to generate a large set of N = 1000 random diffeomorphic 3D surfaces deformations. In the spirit of computational anatomy, for each such movie, we will compute a time dependent diffeomorphic registration of successive frames, and extract an associated detailed strain map. For each pair of movies M1 and M2, after time registration, we will implement diffeomorphic registrations between corresponding key time frames of M1 and M2. This involves the numerical solving of a high dimensional variational calculus problems by innovative fast non-linear optimal control. From all these diffeomorphic registrations, we will extract multiple characteristics of each movie as well as quantitative "similarities" between pairs of movies. At this stage, powerful data mining techniques such as support vector machines and artificial neural networks will become implementable to generate multi-scale clustering as well as classification of shapes deformations. To handle the heavy computing challenges, we will implement highly parallelized computational schemes on remote high power computing resources. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.