The problem of orienting digital images of chest x-rays, which were captured at some multiple of 90 degrees from the true orientation, is a typical pattern classification problem. In this case, the solution to the problem must assign an instance of a digital image to one of four classes, where each class corresponds to one of the four possible orientations. A large number of techniques are available for developing a pattern classifier. Some of these techniques are characterized by independent variables whose values are difficult to relate back to the problem being solved. If a technique is highly sensitive to the values of these variables, the lack of a rigorous way of defining them can be a significant disadvantage to the inexperienced researcher. This thesis presents experiments by the author to solve the chest x-ray orientation problem using four different pattern classification techniques: genetic programming, an artificial neural network trained with back propagation, a probabilistic neural network, and a simple linear classifier. In addition, the author will demonstrate that an understanding of the design of a feature set may allow a programmer to develop a traditional program which does an adequate job of solving the classification problem. Comparisons of the different techniques will be based not only on their success at solving the problem, but also on the time required to find an acceptable solution and the degree to which each technique is sensitive to the values of the variables which characterize it. The thesis demonstrates that all of the techniques can be used to derive very accurate chest x-ray orientation classifiers. While it is dangerous to generalize the results of these experiments to pattern classification problems in general, the author will argue that the magnitude of the differences in performance between the different techniques minimizes this danger. In particular, the experiments suggest that the linear classifier is so computationally inexpensive that it is always worth trying, unless there is a priori knowledge that it will fail. The experiments also suggest that genetic programming is much more computationally expensive than are the linear classifier, artificial neural network, and probabilistic neural network techniques. Of the four conventional pattern classification techniques which were examined, it will be shown that the artificial neural network produced the most accurate classifiers for the x-ray orientation problem. In addition, the results of a number of trials suggest that the final accuracy of the classifier is relatively insensitive to the values of the parameters which characterize this technique, making it an appropriate choice for the inexperienced researcher. With respect to the ability of the resulting classifier to accurately orient sample x-rays which were not included in the training set, the artificial neural network performed well, when compared to the other techniques. Although the classifiers produced by the genetic programming technique were significantly more expensive to construct and were slightly less accurate than the best artificial neural networks, the results of genetic programming experiments can provide insights into the problem being studied, which would be difficult to discern from the classifiers produced by the other techniques. For example, one of the classifiers which was produced by genetic programming uses only eight of the twenty feature values extracted from the sample x-ray. Not only does this reduce the cost of extracting the feature values from an unknown sample, but the classifier itself would be much more efficient to evaluate than the classifiers produced by any of the other techniques.

Library of Congress Subject Headings

Optical pattern recognition; Pattern recognition systems; Chest--Radiography

Publication Date


Document Type


Department, Program, or Center

Computer Science (GCCIS)


Anderson, Peter

Advisor/Committee Member

Gaborski, Roger

Advisor/Committee Member

Biles, John


Note: imported from RIT’s Digital Media Library running on DSpace to RIT Scholar Works in December 2013. Physical copy available through RIT's The Wallace Library at: TA1650 .H644 1997


RIT – Main Campus