Abstract
The present work focuses on spatial/frequency analysis of epiluminesence images of dysplastic nevus and melanoma. A three-level wavelet decomposition was performed on skin-lesion images to obtain coefficients in the wavelet domain. A total of 34 features were obtained by computing ratios of the mean, variance, energy and entropy of the wavelet coefficients along with the mean and standard deviation of image intensity. In order to select features that are statistically correlated, normally distributed features were compared using an unpaired t-test and non-normally distributed features were compared using the Wilcoxon rank-sum test. For our data set, the statistical analysis of features reduced the feature set from 34 to 5 features. For classification, the discriminant functions were computed in the feature space using the Mahanalobis distance. ROC curves were generated and evaluated for false positive fractions from 0.1 to 0.4. Most of the discrimination functions provided a true positive rate for melanoma of 93% with a false positive rate up to 21%.
Original language | English (US) |
---|---|
Pages (from-to) | 1346-1353 |
Number of pages | 8 |
Journal | Proceedings of SPIE - The International Society for Optical Engineering |
Volume | 4684 III |
DOIs | |
State | Published - 2002 |
All Science Journal Classification (ASJC) codes
- Electronic, Optical and Magnetic Materials
- Condensed Matter Physics
- Computer Science Applications
- Applied Mathematics
- Electrical and Electronic Engineering
Keywords
- Dysplastic
- Mahanalobis distance
- Melanoma
- Nevus
- ROC curve
- Wavelet transform