Abstract:
HTML
Purpose: The purpose of this study is to investigate psychophysical measures
in determining quantitatively radiologic similarity for pairs of breast mass
lesions on mammograms to select similar images which can be used as a guide for
radiologists in the distinction of cancer from non-cancer masses.
Methods and Materials: We employed the Digital Mammography Database developed
and distributed via website by the University of South Florida. Regions of
interest (ROIs) of 5 cm by 5 cm in size were selected which included single
mass lesions that could be outlined by a radiologist. A total of 681 malignant
and 763 benign ROIs were used in this study. Contrast and density of mass
lesions were modified to a proper level to facilitate the visual comparison of
the lesions. Five radiologists provided the subjective similarity ratings on a
continuous rating scale between 0 and 1, where 0 expresses as lesions not
similar at all, and 1 as lesions almost identical, for 53 pairs of mass lesions
based on the overall impression for diagnosis of the lesions. We determined 21
image features based on gray-scale distributions around the lesions and 6
morphological features based on the outline of the lesions; the effect of
individual features and their combinations were examined as objective measures.
An artificial neural network (ANN) was employed to learn the relationship
between subjective similarity ratings and objective image features. The ANN was
trained with selected image features as input and average subjective ratings as
output for teaching data. Thus the trained ANN provided a new psychophysical
similarity measure for an unknown pair of mass lesions when the corresponding
set of image features was entered as input. Correlation coefficients between
the subjective ratings and the similarity measures were determined to evaluate
the effectiveness of the similarity measures.
Results: Although the correlation value between subjective ratings and the
combination of features were low (r = 0.515), the correlation value between subjective
ratings and the psychophysical similarity measures obtained by ANN output was
improved (r = 0.682). When only 31 cases with a small variation in subjective
ratings by the five radiologists were employed for training and testing of ANN,
the correlation value was further improved (r = 0.750).
Conclusion: A new psychophysical measure for the similarity of mass lesions
would be useful for selecting similar mass lesions, which would assist
radiologists in distinguishing benign from malignant mass lesions on
mammograms.
(R.A.S. is a shareholder in R2 Technology Inc. K.D. is a shareholder in R2
Technology Inc. and Deus Technology Inc.)
Questions about this event email: chisa@uchicago.edu
Muramatsu, C,
Usefulness of Psychophysical Measures for Selection of Similar Images for Distinction between Benign and Malignant Mass Lesions on Mammograms: A Pilot Study. Radiological Society of North America 2003 Scientific Assembly and Annual Meeting, November 30 - December 5, 2003 ,Chicago IL.
http://archive.rsna.org/2003/3103006.html
