Pattern recognition in medical imaging

Project: Research project

Description

Recent work in pattern recognition has demonstrated that computers can equal or even surpass image classification and pattern analysis by human experts. Modern imaging systems far exceed the human eye in spatial and spectral resolution as well as dynamic range, thus potentially allowing machine-based image pattern analysis systems to perform such tasks. The pattern analysis system we have developed, characterized and published is called WND-CHARM. The approach is based on extracting over 2,000 descriptors of image content from each image using both standard feature extraction algorithms as well as others we developed. A key innovation was to extract image content both from the original images, and also from image transforms such as Fourier and wavelet. Each descriptor is assigned a weight based on its ability to discriminate between training sets of images. Each training set ("class") contributes to the marginal probability that a previously unseen test image belongs to one of the defined classes. A key property of this classification mechanism is that marginal probabilities can be interpreted as image similarities, thus producing quantitative measures of similarity rather than merely qualitative classifications. A substantial effort this year was devoted to applying this technique to the study of sarcopenia (age-related muscle degeneration) in the worm C. elegans. We were able to discern three stable morphologies and transitions between them during the adult worm lifespan. This is the first demonstration of stable post-developmental morphological states in any organism. The implications are that aging is not a fully stochastic process, and that it may be possible to substantially alter the aging process by interfering with these transitions. The generality of our classification system has allowed us to investigate its potential for medical diagnosis as well. In an on going study of lymphoma diagnosis and classification in a clinical setting using histological sections from human biopsies stained with Hematoxylin/Eosin (H/E), the variation in sample preparation in a small random collection of existing slides was too great to adequately perform diagnosis (i.e. determine the lymphoma class for an entirely new case); but our algorithms could differentiate the three lymphoma classes more accurately than trained pathologists. A significant contribution of this work was the comparison of several approaches to pattern recognition in histological H/E sections, and identification of critical parameters necessary for machine-based diagnosis. This work has been recently submitted for publication. Our work with histological H/E sections will continue this year in a collaboration with Ashani Weeraratna to study melanoma. Our initial results indicate that carefully controlled sample preparation and staining are adequate to differentiate melanomas from nevi and also to accurately differentiate secondary tumor sites by location. The goal of this study is to find early structural biomarkers that would predict the aggressiveness of the primary tumor. Our approach to pattern-recognition is not inherently restricted to images obtained by microscopy. We investigated its application to the diagnosis of osteoarthritis (OA) in human knee X-rays obtained from the Baltimore Longitudinal Study of Aging (BLSA). Our published work shows that we can diagnose both moderate (KL grade 3) and minimal (KL grade 2) OA with an accuracy of 91.5% and 80.5% respectively. The reference diagnosis we used for each X-ray image was made by two independent radiologists, with a third radiologist adjudicating discrepancies, making the reference diagnosis much more accurate than one from a single radiologist. We are currently making use of longitudinal BLSA data to investigate if our pattern recognition technique can be used to predict the occurrence of OA in subsequent visits. Planned studies for this year include applying this technique to MRI images of ostoarthritis.
StatusFinished
Effective start/end date4/1/963/31/02

Funding

  • National Institutes of Health: $107,784.00
  • National Institutes of Health: $81,810.00
  • National Institutes of Health: $107,784.00
  • National Institutes of Health
  • National Institutes of Health: $93,420.00

Fingerprint

Adrenergic Receptors
Weight Loss
Fats
Adipocytes
Cardiovascular Diseases
Intra-Abdominal Fat
Abdominal Obesity
Receptors, Adrenergic, beta
Body Fat Distribution
Lipolysis
Metabolic Diseases
Adiposity
Oils and fats
GTP-Binding Proteins
Immunoblotting
Metabolism
Radioimmunoassay
Adipose Tissue
Obesity
Cell Membrane

ASJC

  • Medicine(all)