Development of new descriptor for melanoma detection on dermoscopic images.

Early detection of melanoma has critical importance for the success of the treatment. However, a successful early diagnosis is only possible with the existence of discriminative features. In this study, a new descriptor based on the number of colors was developed in order to successfully diagnose lesions of melanoma. The number of colors is the main feature in the identification of melanoma-type skin lesions. The user must select a threshold value when calculating the number of colors of the lesion. The incorrect threshold value selection of non-expert users disrupts the aforementioned feature and also leads to significant diagnostic errors. In this study, it was revealed that color counting threshold values have a significant effect on the distinctiveness of the number of colors. In the three dermoscopic databases, color counting threshold values that provide the maximum distinctiveness on melanoma and benign lesions were determined as 0 and 0.123 respectively. By using these color counting threshold values, the number of colors for each sample in the data sets was calculated separately. Following that, a novel attribute called the number of color difference was defined as a function of color counting threshold values. Experiments using only the proposed new descriptor yielded 52.7% higher f-measure and 84.5% higher true-positive performance than the number of colors used in the literature. The results obtained in this study revealed the importance of accurately determining the number of colors the lesions had and states that the applied color counting threshold significantly influences the classification results. Thereby, a new method is proposed for determining the critical color counting threshold. We claim that the classical ABCD rule should be improved by our new descriptor. Graphical abstract Fig. 1 Selection of threshold has vital effect on skin lesion classification. A new method to select the correct threshold value and a new attribute for correct classification were developed.