Fractals for Malignancy Detection in Dermoscopy Images

Accurate diagnosis of melanocytic lesions is amongst the most difficult problems for dermatologists. Border irregularity of a skin lesion is one of the diagnostic criteria to be assessed by the dermatologist. While there are myriad publications defining the dermatologic criteria that reproducibly distinguish ("presumably") benign melanocytic nevi from malignant melanomas, these criteria are neither universally accepted nor easily recognizable in all cases. To close this gap, this study focuses on quantitative assessment of shape-based irregularity features of suspected skin lesions in dermoscopy images. Border irregularities were investigated and analyzed in 100 skin lesions to develop objective and quantifiable criteria that evaluate diagnostically challenging lesions and effectively distinguish benign from malignant lesions. More specifically, this study automatically delineates skin lesion borders and then quantitatively measures irregularity of the extracted border by using eleven different fractal measures. Classification and feature selection analysis showed that malignancy in dermoscopy images can be detected with a high accuracy by using fractal features. We also find that which fractal features are the most effective ones for finding malignancies.