CCBM Seminar Series

Albert Montillo

Doctoral Candidate, University of Pennsylvania

Developing biomedical imaging and modeling as tools for understanding function and structure

In this talk I describe the methods we have developed through my Ph.D. research for cardiac image analysis as well as some methods in brain image analysis.      

    We begin by adopting a rigorous statistical analysis framework, in which we systematically rank several image processing strategies. Then using principles from the physics of MRI, we devise a strategy for handling both thermal noise and RF inhomogeneity which has applicability for improving image quality for many organs including the heart and brain.

      Focusing on cardiac image analysis, we develop automated segmentation and analysis methods for the left and right ventricles in SPAMM-MRI. This pulse sequence had been developed to provide in-vivo measurement of regional heterogeneity of myocardial contraction and potentially reduce cardiovascular disease morbidity. However, widespread adoption of the pulse sequence has been hindered by the inordinate amount of interactive analysis time. Typically 80% of the time is spent extracting the endocardial and epicardial ventricular surfaces. Previous research left this step largely manual due to the many challenges an automated system would face. To automate this step, we have developed a method which is based on (1) a sequence of image processing steps and (2) a finite element, model-based segmentation framework. We consistently locate the heart using a novel spatiotemporal pruning algorithm, which is also suitable for new SSFP pulse sequences. To segment the ventricles, we derive 3D forces from 2D images to fit a bi-ventricular heart model to patient data. We perform both qualitative and quantitative evaluation of our segmentation for normal and diseased subjects. The surfaces we segment are inherently valuable 1) to measure classical descriptors of cardiac function 2) to provide boundary conditions for a biomechanical PDE model and 3) to improve the accuracy of tag tracking. I conclude this section with results from our methods for the analysis of myocardial deformation between theses surfaces.

     On brain imaging, I will describe our method to automate the labeling of neuroanatomical structures in MR brain volumes. While most existing procedures label only a small number of tissues, our method assigns one of 37 labels to each voxel and has been found to yield measurements which are statistically indistinguishable from manual labeling and have sufficient sensitivity to detect changes that presage the onset of probable Alzheimer's disease.

For information on disability access, contact Anne Albinak at 410-516-5310 or aalbinak@bme.jhu.edu

Wednesday, February 18, 2004
1:00-2:00pm

Room 210,
Clark Hall