University of Minnesota
University Relations

Minnesota Supercomputing Institute

Log out of MyMSI

Research Abstracts Online
January 2009 - March 2010

Main TOC ....... College TOC ....... Next Abstract

University of Minnesota Twin Cities
Institute of Technology
Department of Biomedical Engineering

PI: Narendra K. Simha
Co-PI: Ahmed Tewfik

Segmentation Methods for Quantitation MRI of Musculoskeletal Tissues

MRI segmentation of human cartilage and bone in the knee joint is relevant for the diagnosis of osteoarthritis (OA) and for evaluating OA progression and joint repair, as well as for joint biomechanical models to design prostheses, study motion disorders, or optimize surgical methods. Delineation of cartilage in the human knee is particularly challenging because cartilage occupies only about 5% of the knee joint and various joint tissues and adjacent muscle have the same MRI intensities as cartilage. Consequently, there has been little success with automated methods and most reports in the literature are semi-automated at best and invariably use human input. Professor Simha’s group has developed a Mimics-based segmentation method to extract solid models of human knee cartilage and bone. Although the approach is faster than published values, it still requires about an hour. Hence there is a critical need for faster and preferably automatic methods to segment the knee cartilage and bone from MR images. Professor Tewfik’s group has recently developed a method to reconstruct deformable three-dimensional organs, called sparse parametric representation of deformable objects (SPRDO). It has substantially better computational speed and lower memory requirements than existing algorithms. Organs are represented by parameters in the frequency domain, by using a spherical harmonics transformation. Then a dictionary learning method is applied to a large set of training models to identify a sparse parameter subset. The ability of this principal basis to reconstruct the organ is determined using a different set of evaluation models.   

The goal of this project is to develop automatic segmentation methods to extract solid models of the human knee from MR images through the following tasks: evaluate the ability of the SPRDO method to reconstruct human knee cartilage and bone shape; and extend the SPRDO method to consider image intensity and texture and thereby segment human knee cartilage and bone from MR images.

Group Members

Sundareswaran Kapaleeswaran, Graduate Student
Dan Wang, Graduate Student