Medical Abstract
COMPUTER ASSISTED CLASSIFICATION OF SKIN PIGMENTED LESIONS
Rosamaria Corona, M.D.; Istituto Dermopatico dell'Immacolata, Via dei monti di Creta 104, Rome, Italy; Marco Burroni, Ph.D.; Lega Italiana Tumori, Siena, Italy; Pietro Puddu, M.D.; Istituto Dermopatico dell'Immacolata, Rome, Italy; Lucio Andreassi, M.D.; Department of Dermatology, University of Siena, Siena, Italy
American Academy of Dermatology 60th Annual Meeting
Feb 22, 2002 to Feb 27, 2002
Poster Discussion
Abstract P10
Key Opinion Leader Commentary
Submitted by: Mark Beckett, MD on Mar 23, 2006
[Commentary] The linear classifier was further tested on a set of 16,069 lesions, observed in several Italian centers using the same instrumentation, and judged as benign lesions for which excision was not warranted. The specificity obtained ranged from 73% to 92%. The feasibility of automatic differentiation of melanoma from benign melanocytic lesions is supported by this study. The stability of the classifier performance across different test sets is likely to reflect objective measurable morphological differences between melanomas and benign lesions, although further verification on different lesion populations is needed.
The aim of this study was to evaluate the performance of a linear classifier, trained for sensitivity close to 100% and high specificity, on a set of histologically confirmed melanomas and benign pigmented lesions observed at the Istituto Dermopatico dell'Immacolata, Rome, Italy, and at the University Dermatology Department, Siena, Italy. Images were taken, stored and processed using the DBDermo-Mips digital dermoscopy system, which automatically evaluates the border of the lesion and 49 quantitative variables related to lesion morphology. To build the linear classifier, after a preliminary analysis performed on all parameters, a subset of 14 not highly correlated variables was selected, and a linear combination of these independent variables was computed by discriminant analysis. For each lesion a score was obtained and a threshold value identified that allowed the assignment of each melanoma to the correct group with a sensitivity of 95%. The linear classifier was trained on a random sample of 184 melanomas and 220 benign melanocytic lesions, extracted from the Rome and Siena pooled image databases.
The classifier performance was assessed using as test set 78 melanomas and 92 nevi observed at the Rome center and 110 melanomas and 137 nevi observed at the Siena center, and yielded a sensitivity of 95% and a specificity of 88% for the IDI test set, and a sensitivity of 93% and a specificity of 77% for the Siena test set.
The linear classifier was further tested on a set of 16,069 lesions, observed in several Italian centers using the same instrumentation, and judged as benign lesions for which excision was not warranted. The specificity obtained ranged from 73% to 92%. The feasibility of automatic differentiation of melanoma from benign melanocytic lesions is supported by this study. The stability of the classifier performance across different test sets is likely to reflect objective measurable morphological differences between melanomas and benign lesions, although further verification on different lesion populations is needed.