Tulips to Thresholds
John W. Swets
April 12, 2010
April 12, 2010
Reviewed by Charles Metz (published in the December 2010 issue of Academic Radiology) To see full review please go to Charles Metz's Review
Although receiver operating characteristic (ROC) methodology has become ubiquitous in evaluation of medical imaging technology, this was not always so; in fact, the fundamental insight that detection or classification performance can be described most meaningfully in terms of a variable tradeoff between true-positive and false-positive rates did not emerge in any field until around 1950. John Swets was there at the beginning, initially as a graduate student in psychology at the University of Michigan and subsequently during an illustrious career that included substantial research and managerial accomplishments in academic, corporate and public settings. As its subtitle indicates, this memoir traces the parallel paths of Dr. Swets’ career and development of signal detection theory, which provides the conceptual basis upon which ROC analysis has been built.
The Tulips of the book’s title refers to Dr. Swets’ Dutch heritage and early life in Holland, Michigan, whereas the plural Thresholds harks to his (and our) conceptual transition from one kind of “threshold” to another. Until the mid-twentieth century, human perception was thought to involve a fixed threshold such that no sensory input below some minimum level of stimulation produced a corresponding sensation; thus, false-positive detections were ascribed to malfeasance or hallucination. Signal detection theory (SDT) posited, instead, that sensory “signals” must be detected within a context of underlying “noise” — inherent statistical variation — that could both mask real signals and mimic signals when none was present. From this perspective, a signal is “detected” or a difference between two kinds of signals is “recognized” by comparing the statistically-variable stimulus with a variable decision threshold that the human observer (or a machine observer) can, in principle, adjust to accommodate different signal prevalences and/or different costs of false-positive and false-negative errors, etc. To make a decade-long story short: SDT won out and now provides both the conceptual and the quantitative basis for formally making and evaluating most decisions that involve statistical uncertainty.
more ...
Although receiver operating characteristic (ROC) methodology has become ubiquitous in evaluation of medical imaging technology, this was not always so; in fact, the fundamental insight that detection or classification performance can be described most meaningfully in terms of a variable tradeoff between true-positive and false-positive rates did not emerge in any field until around 1950. John Swets was there at the beginning, initially as a graduate student in psychology at the University of Michigan and subsequently during an illustrious career that included substantial research and managerial accomplishments in academic, corporate and public settings. As its subtitle indicates, this memoir traces the parallel paths of Dr. Swets’ career and development of signal detection theory, which provides the conceptual basis upon which ROC analysis has been built.
The Tulips of the book’s title refers to Dr. Swets’ Dutch heritage and early life in Holland, Michigan, whereas the plural Thresholds harks to his (and our) conceptual transition from one kind of “threshold” to another. Until the mid-twentieth century, human perception was thought to involve a fixed threshold such that no sensory input below some minimum level of stimulation produced a corresponding sensation; thus, false-positive detections were ascribed to malfeasance or hallucination. Signal detection theory (SDT) posited, instead, that sensory “signals” must be detected within a context of underlying “noise” — inherent statistical variation — that could both mask real signals and mimic signals when none was present. From this perspective, a signal is “detected” or a difference between two kinds of signals is “recognized” by comparing the statistically-variable stimulus with a variable decision threshold that the human observer (or a machine observer) can, in principle, adjust to accommodate different signal prevalences and/or different costs of false-positive and false-negative errors, etc. To make a decade-long story short: SDT won out and now provides both the conceptual and the quantitative basis for formally making and evaluating most decisions that involve statistical uncertainty.
more ...