Abstract:
HTML
Purpose: To assess the benefit of prevalence modification to the performance
of radiologists using multi-modality intelligent workstations for the diagnosis
of breast lesions.
Methods and Materials: Our current multi-modality intelligent workstation utilizes
two estimates of a lesion's probability of malignancy (PM): one obtained from a
classifier trained on a sonographic database of 356 cases (cancer
prevalence=0.19) and an another from a classifier trained on a mammographic
database of 267 cases (cancer prevalence=0.55). These estimates of the PM are
dependent on the prevalence of cancer in each training database, which usually
does not correspond to the prevalence best suited to the user, for example, the
prevalence of the population seen in the user's medical practice. We use an
estimate of the prevalence best suited to the user to transform the
computer-estimated PM into a prevalence-modified estimate of the PM by an
application of Bayes' rule. An estimate of the prevalence best suited to the
user is determined by relating the radiologist's PM to that of the
computer-estimated PM for a given set of breast lesions.
Results: A radiologist gave a PM for each case in a mammographic and
sonographic database (97 lesions in each case) and these values of the PM were
used to estimate the radiologist's operating prevalence by determining the
least squares fit to a Bayes' rule transformation of the computer-estimated PM.
The radiologist's operating prevalence for the mammographic database was 0.46
and for the sonographic database was 0.51. Before prevalence modification, the
correlation between the radiologist PM and the computer PM was 0.48 and 0.47
for the mammographic and sonographic databases, respectively. After prevalence
modification, the correlation was 0.73 and 0.75 for the mammographic and
sonographic databases, respectively.
Conclusion: Our preliminary study indicates that prevalence modification
increases the correlation between the radiologist-estimated probabilities of
malignancy and the computer-estimated probabilities of malignancy. It is
expected that intelligent workstations that use prevalence-modified
probabilities of malignancy will improve the performance of radiologists in the
task of differentiating malignant and benign breast lesions. (M.L.G., C.J.V., C.M.,
G.N., and R.A.S. are shareholders in R2
Technology, Inc.)
Horsch PhD, K,
Prevalence-modified Estimation of Computer-determined Probabilities of Malignancy for CAD. Radiological Society of North America 2003 Scientific Assembly and Annual Meeting, November 30 - December 5, 2003 ,Chicago IL.
http://archive.rsna.org/2003/3107305.html