Computer Diagnosis of Primary Bone Tumors

Abstract

The accuracy of diagnosis is inevitably related to the consideration of all possible disease entities which may explain a set of symptoms and findings. Statistical weighing of each of these possibilities in the light of the established symptoms and findings is necessary to determine which one has the greatest probability of being correct. Physicians have been performing these mental gymnastics with varying degrees of effectiveness since man assumed responsibility for the medical care of his kind. When the range of diagnostic possibilities or the number of recognized manifestations increases to such an extent that they can no longer be handled efficiently, physicians respond by specialization and in this way limit the number of diseases to a range which can be comfortably managed. We are fast approaching the limit to which we can further specialize and yet continue to meet the needs of society, even though the number of new facts about disease and methods for patient care are being increased at an unprecedented rate. The electronic digital computer, first developed and used only sixteen years ago, has experienced an explosive rate of application in other branches of science and in industry, because of its unique and indispensable capacity to retain large numbers of facts and to accept an exact and detailed program of instructions describing how to interrelate these facts in order to provide a statistically weighed answer. Awareness of the usefulness of such an instrument in medical practice is rapidly growing. This paper reports our progress in the development of a program of instructions to enlist the aid of a computer in the evaluation of cancer of bone as visualized on roentgenograms. Requisites for Programming Diagnosis There are at least five essentials in the development of a program of instructions for use by a computer as an aid to diagnosis. These are as follows: 1. We must make the assumption that the proper diagnosis can actually be made in a large percentage of cases. Some time ago, one of the authors (G.S.L.), in a study of a large series of tumors, predicted the most likely histologic diagnosis for each case by examination of the roentgenograms. He found that his average rate of accuracy in predicting the correct histologic type was 80 per cent, with benign chondroblastoma and osteosarcoma being most diagnosable, and reticulum-cell sarcoma being least so. Subsequent experience with individual cases in a clinical setting has shown improvement in this rate, particularly in respect to reticulum-cell sarcoma. One factor quite likely to reduce the overall success of predicting histologic types by the use of data from roentgenograms is the variability of opinion among pathologists. While pathologists as a group adhere to well defined histologic classifications of bone tumors, in practice one finds considerable differences of opinion, and even disagreement as to the diagnosis in an individual case.