Journal of Testing and Evaluation | 2021
Fuzzy Ontology for Automatic Skin Lesion Classification
Abstract
Medical diagnosis of cancer is becoming more complex in recent years, with doctors focusing on making both earlier and accurate diagnoses to save patients’ lives. Such goals are more challenging for melanoma, which is the deadliest of skin cancers. Recently, melanoma incidence has increased significantly because of climate change. Fortunately, early detection leads to a 5-year survival rate of 98 %. Computer-aided diagnosis systems can offer a more objective analysis tool, taking into consideration the expert’s knowledge. Ontology offers an efficient framework for reducing the gap between low-level information and expert analysis. A dermatologist’s recommendation is often based on the ABCD rule, involving four characteristics of a lesion, which are asymmetry, border, color, and differential structures. A score, associated to a qualitative description of the lesion, allows lesions to be categorized into three classes: melanoma, benign, or recommended follow-up. Early research on automatic diagnosis doesn’t take into consideration either the qualitative aspect of expert description or the ambiguity of information itself. In this paper, we propose a fuzzy ontology–based melanoma diagnosis system. A fuzzy classifier is proposed to cope with the qualitative description of experts. Then, a fuzzy inference system is proposed to generate the decision. Experimental validation is undertaken on both optical and dermoscopic images from public datasets DermQuest, Dermatology Information System, and International Skin Imaging Collaboration (ISIC). For optical images, we get a sensitivity of 91 %, a specificity of 88 %, and an accuracy of 90 %, whereas for dermoscopic images, we obtain a sensitivity of 92 % and 91 %, a specificity of 91 % and 93 %, and an accuracy of 91 % and 92 % for ISIC 2016 and ISIC 2017, respectively. A comparative study with existing approaches shows that these performances ensure higher accuracy rates and the best compromise between sensitivity and specificity.