NEW METHODS OF PATTERN RECOGNITION IN MEDICAL THERMOGRAPHIC SCREENING

KENNETH JAMES SCHLAGER, Marquette University

Abstract

Medical thermography is the pictorial or computerized image representation of multiple, simultaneous temperature measurements of the skin surface of the human body for the purpose of medical screening and diagnosis. Its primary application has been in breast cancer detection. Breast cancer is the leading cause of death by cancer in women with over 35,000 deaths annually in the United States alone. One out of every 15 women will at some time develop breast cancer and in spite of new innovations in surgery, radiotherapy, chemotherapy and diagnosis, the death rate has remained essentially constant at 22 per 100,000 female population for the past 40 years. The five year survival rate depends on the size of the cancer and the metastatic spreading existing at the time of diagnosis. Cure rates exceed 90% if the cancer is detected at an early stage of development. As the cancer grows, however, the cure rate declines to about 43% in the later stages. This relationship between early detection and cure rate has stimulated research and development in breast cancer screening and diagnostic techniques. A research study was undertaken to determine whether syntactic as opposed to statistical methods of computer-based pattern recognition could improve the diagnostic accuracy of medical thermography as applied to breast cancer detection. Statistical pattern recognition emphasizes the use of image "features" embodied in a linear function with the parameters derived from statistical analyses of a data base. Syntactic pattern recognition provides for direct comparison of images and emphasizes the structural characteristics of an image. A special instrumentation system using an 8 x 8 array of infrared sensors and a microcomputer was designed and constructed to provide a vehicle for gathering the clinical test data needed to support this research project. Clinical tests of 641 women were conducted in a breast cancer detection screening center at the Medical College of Wisconsin in 1979 and 1980. This clinical data base was analyzed with two significantly different methods of pattern recognition. The first method, linear discriminant analysis, a form of statistical pattern recognition, was applied using parameters derived from an earlier data base from Drexel-Temple in Philadelphia. Five different linear discriminant functions were used to provide a comprehensive evaluation of this technique of pattern recognition. Three of these functions were newly derived for this research, and two were discriminant functions previously used by Negin at Drexel-Temple. The screening test results of this evaluation were similar to those experienced at Drexel-Temple and the University of Oklahoma indicating the general validity of the experiment. These two data bases were then used to evaluate the new syntactic method of pattern recognition that emphasized recognition of the structural characteristics of the thermal patterns. The Drexel data base was used to synthesize a set of reference images known as templates which were matched against the thermal images of the patients in the Milwaukee data base. This new method of pattern recognition, syntactic template matching, demonstrated superior sensitivity and specificity in the clinical test evaluation. Sensitivity refers to the percentage of true positives diagnosed correctly while specificity designates the percentage of true negatives detected by the system. A sensitivity of 75.0% was demonstrated along with a specificity of 95.2% by the syntactic template matching method. This performance compared with a sensitivity of 65.0% and a specificity of 76.7% for the best discriminant analysis function.

Recommended Citation

KENNETH JAMES SCHLAGER, "NEW METHODS OF PATTERN RECOGNITION IN MEDICAL THERMOGRAPHIC SCREENING" (January 1, 1980). Dissertations (1962 - 2010) Access via Proquest Digital Dissertations. Paper AAI8111867.
http://epublications.marquette.edu/dissertations/AAI8111867

Share

COinS