Owen Carmichael : Research : Medical

Home / News Affiliations / Contact Research Medical Image Analysis Object Recognition From Images Object Recognition From Range Data 3D Modeling From Range Data Super-Resolution Image Retrieval Papers Teaching Community Resources Reading Groups Leisure

Variability in progression of Alzheimer’s disease Alzheimer’s disease (AD) is characterized by an insidious loss of cognitive function over time, but the rate of cognitive decline varies widely from person to person. I am currently working on trying to understand the variability in cognitive decline rates; specifically, I am looking at relationships between rates of cognitive decline and regions of interest in the brain. I am using fully-automated image processing techniques (see below) to isolate the regions of interest in structural MR images. For one study, we used fully-automated segmentation techniques to isolate the lateral ventricles in over 300 subjects who received MR images in 1998 as part of the Cardiovascular Health Study. The subjects received clinical evaluations in 1998 and 2002, and were placed into categories according to their rates of cognitive decline between 1998 and 2002. Among subjects who were normal in 1998, we found that those who rapidly declined to dementia by 2002 (represented by the magenta curve below) had larger lateral ventricles in 1998 than subjects who stayed normal (red curve) or subjects who declined to a transitional state known as mild cognitive impairment (MCI, orange curve). We hope that by understanding relationships between cognitive decline rates and brain structure characteristics, we could be able to predict and treat cognitive decline more effectively. Web site: The Cardiovascular Health Study Fully-automated brain segmentation During my postdoc, I focused on automatically localizing two of these parts of the brain—the hippocampus and the lateral ventricles. I completed an extensive empirical study of many competing methods for this problem, all of which are unified by the general theme of aligning a patient’s image with a reference image on which the hippocampus has been traced by an expert. One way to distinguish the different methods is by the degree to which they geometrically deform the patient image while aligning it to the reference image. This figure shows how three different techniques deform the patient image: Note that some of the methods (like “Chen Fully-deformable”) deform the image in a more complicated, irregular way, while some of the other ones (like “AIR Semi-deformable”) deform the image more smoothly and gradually. Our experiments have shown that techniques that use the more complicated deformations tend to do a better job at localizing the hippocampus. Paper: O. Carmichael, H. J. Aizenstein, S. W. Davis, J. T. Becker, P. M. Thompson, C. C. Meltzer, Y. Liu. Atlas-Based Hippocampus Segmentation In Alzheimer’s Disease and Mild Cognitive Impairment. Carnegie Mellon University Robotics Institute Technical Report CMU-RI-TR-04-53, December 2004. [html] Web site: Pittsburgh ADRC, my research center. It turns out that these irregular, fully-deformable alignment techniques are useful for other medical image analysis problems related to geriatric imaging. I have been helping Dr. Michael Gach and his grad student, Weiying Dai, use fully-deformable alignment techniques to align a set of elderly structural/functional images to a common coordinate frame so that atrophy effects are cancelled out. They are using the approach to analyze continuous arterial spin labeling (CASL) images of populations of elderly patients with AD and Mild Cognitive Impairment (MCI). Paper: Weiying Dai; Oscar L. Lopez; Owen T. Carmichael; James T. Becker; Lewis H. Kuller; and H. Michael Gach Abnormal Regional Cerebral Blood Flow in Cognitively Normal Elderly Subjects With Hypertension. Stroke, In Press. [link]. Web site: Michael Gach’s page Also, I have been helping Dr. Howard Aizenstein and his grad student, Minjie Wu, use fully-deformable alignment to localize structures and do group-wise analysis in functional MRI (fMRI) images of depressed elderly subjects doing learning tasks. Paper: M. Wu, O. Carmichael, C. S. Carter, J. L. Figurski, P. Lopez-Garcia, H. J. Aizenstein. (2005, In Press). Quantitative comparison of neuroimage registration by air, spm, and a fully deformable model. Human Brain Mapping. Web site: Howard Aizenstein’s web page Fully-deformable image alignment is helpful for a variety of problems, but the problem of how to estimate the deformation is important. That is, what information do you pull out of the two images that gives you a good guess at how to line them up? I have been helping to advise Leonid Teverovskiy, Dr. Yanxi Liu’s grad student, on a technique for automatically determining which aspects of images (which features) should be used as cues for how to align them. Also, we have worked on automatically determining which features are useful for automatically classifying whether images correspond to healthy subjects or those with MCI or AD. Web site: Yanxi Liu Paper: Y. Liu, L. Teverovskiy, O. Carmichael, R. Kikinis, M. Shenton, C. Carter, A. Stenger, S. Davis, H. Aizenstein, J. Baker. Discriminative MR Image Feature Analysis for Automatic Schizophrenia and Alzheimer's Disease Classification. CMU RI Technical Report CMU-RI-TR-04-15 [html]. [top]