The concepts of relative operating characteristics (ROC-curves) and detectability index (d') are introduced for the purpose of evaluating performance in breast thermography. In assessing published information on the subject we have found that much of it is anecdotal and lacks sufficient data to determine performance. We have also found that for those published findings which had sufficient data i t was possible to reconcile conflicting conclusions as to the efficacy of thermography, and that performance could be quantitated with the use of the detectability index. Conclusions for optimizing available clinical thermographic techniques are given together with ways for future improvement.Cancer 36: 2 159-2 163, 19 75.
REAST T H E R M O G R A P H Y AS A CLINICAL TOOL ISB now almost 20 years old. While its potential as a clinical screening tool is recognized by many 2*6*11*13 there exists also a significant number of those who question the technique's usefulThe reason for disagreement on the value of breast thermography as a screening technique seems to arise out of disagreement in the answers given to questions concerning: 1) the technique's specificity and sensitivity (i.e. the false-negative and false-positive rates); 2 ) the consistency of criteria used (i.e. the ability to arrive at the same decision on a given case by different observers or by the same observer at repeated viewing); and 3 ) the weight ascribed to thermographic findings when used in conjunction with other examination modes.In this paper we will attempt to bring various data reported in the literature to a common denominator, evaluate their significance and consider methods for improvement.
Specificity and Sensitivity of Breast Thermo-In surveying available data on the ability of thermograms to separate normal from abnormal (or benign from malignant) cases one is struck P P h Y by the wealth of anecdotal information, i.e. descriptions of cases where thermography revealed malignancies undetectable by other diagnostic means. At the same time the scarcity of good statistical data is also quite evident. The latter are necessary to prove that the technique can indeed provide a useable tool for diagnostic purposes. For these purposes two figures are needed: the false-negative rate and the false-positive rate. These two types of error rates can be evaluated by the use of the relative operating characteristics (''ROC-curves") described by Swets and ~t h e r s . '~'~~ Such a set of curves is shown in Fig 1. where true-positive decisions are plotted against false-positive decisions. Curves A and B represent two different decision techniques, while points along any one curve represent different observer attitudes (over-or underreading). By changing his threshold criterion or attitude the observer will move along the same curve trading an improvement in true-positive finding for increased false positives and vice versa. His overall accuracy, however, remains unchanged. This performance accuracy can be defined quantitatively by the index of detectability d' t...
