USING MACHINE LEARNING TO CLASSIFY IMAGE FEATURES FROM CANINE PELVIC RADIOGRAPHS: EVALUATION OF PARTIAL LEAST SQUARES DISCRIMINANT ANALYSIS AND ARTIFICIAL NEURAL NETWORK MODELS

Authors

  • Fintan J. McEvoy,

    Corresponding author
    • Department of Veterinary Clinical and Animal Sciences, Faculty of Health and Medical Sciences, University of Copenhagen, Copenhagen, Denmark
    Search for more papers by this author
  • José M. Amigo

    1. Department of Food Science, Quality and Technology, Faculty of Science, University of Copenhagen, Copenhagen, Denmark
    Search for more papers by this author

  • Part of this study was presented in abstract form at the 16th International Veterinary Radiology Association meeting, Bursa, Turkey, 2012.

  • Funding source: Chemometric Analysis Center (CHANCE), University of Copenhagen, Copenhagen, Denmark.

Address correspondence and reprint requests to Fintan J. McEvoy, at the above address. E-mail: fme@sund.ku.dk

Abstract

As the number of images per study increases in the field of veterinary radiology, there is a growing need for computer-assisted diagnosis techniques. The purpose of this study was to evaluate two machine learning statistical models for automatically identifying image regions that contain the canine hip joint on ventrodorsal pelvis radiographs. A training set of images (120 of the hip and 80 from other regions) was used to train a linear partial least squares discriminant analysis (PLS-DA) model and a nonlinear artificial neural network (ANN) model to classify hip images. Performance of the models was assessed using a separate test image set (36 containing hips and 20 from other areas). Partial least squares discriminant analysis model achieved a classification error, sensitivity, and specificity of 6.7%, 100%, and 89%, respectively. The corresponding values for the ANN model were 8.9%, 86%, and 100%. Findings indicated that statistical classification of veterinary images is feasible and has the potential for grouping and classifying images or image features, especially when a large number of well-classified images are available for model training.

Ancillary