New Accurate Automated Melanoma Diagnosing Systems

In this paper we developed two automatic skin lesion classification systems using shape, texture, and color features. 10 lesion features (LFs) were extracted for each sample of a data set of 102 (52 malignant, 50 benign) dermoscopic lesion images. Next, three different rankings of the 10 LFs generated the most significant subsets that improved the melanoma diagnosis. A support vector machine (SVM) classifier found a 6LFs subset with the highest f-measure for melanoma diagnosis. The 6LFs represents a new melanoma diagnosing system. A new approach for melanoma dot and globule quantification expanded the extracted LFs to 11. Of the 11 LFs were generated three rankings, and the most significant subset was extracted producing the highest f-measure for melanoma diagnosis. This designed method represents a second melanoma diagnosing system with 5LFs, including dots and globules. Model, leave one out (LOO), and 10-fold cross validation (10xCV) statistics are used to calculate the f-measure of the two diagnosing systems. A comparison analysis shows that the 5LFs system possesses higher accuracy and sensitivity of melanoma diagnosis than the 6LFs system, which has a higher sensitivity and accuracy than the major clinical and contemporary automated systems.