Melanoma is one of the most dangerous types of skin cancer that its early detection can save patients' lives. Computer‐aided methods can be used for this early detection with acceptable performance. In this study, a system is proposed to detect melanoma automatically using an ensemble approach, including convolutional neural networks (CNNs) and image texture feature extraction. Two CNN models, a proposed network and the VGG‐19, were employed to classify images in the CNN phase. Furthermore, texture features were extracted, and their dimension was reduced using kernel principal component analysis (kPCA) to improve the classification performance in the feature extraction‐based phase. The results of each step were then combined to obtain the final diagnosis. The proposed method was evaluated on three databases, that is, ISIC 2016, ISIC 2019, and PH². The accuracy, average precision, sensitivity, and specificity of the proposed method on the ISIC 2016 dataset were 85.2%, 66%, 52%, and 93.4%, respectively. These evaluation metrics for the ISIC 2019 database were obtained equal to 96.7%, 95.1%, 96.3%, and 97.1%, respectively. Furthermore, the accuracy, sensitivity, and specificity of the proposed method on the PH² dataset were 97.5%, 100%, and 96.88%, respectively. According to the experimental results, the ensemble method improves the evaluation metrics compared to each phase separately. Besides, the proposed approach can increase the performance of melanoma detection, compared to previous studies.

中文翻译：

---

结合卷积神经网络和纹理特征在皮肤镜图像中自动检测皮肤癌

黑色素瘤是最危险的皮肤癌类型之一，其早期发现可以挽救患者的生命。可以使用计算机辅助方法进行早期检测，并获得可接受的性能。在这项研究中，提出了一种使用集成方法自动检测黑色素瘤的系统，包括卷积神经网络（CNN）和图像纹理特征提取。使用两个CNN模型（一个建议的网络和VGG-19）对CNN阶段中的图像进行分类。此外，提取了纹理特征，并使用内核主成分分析（kPCA）降低了其特征的维数，以提高基于特征提取阶段的分类性能。然后将每个步骤的结果合并以获得最终诊断。在三个数据库（即ISIC 2016，ISIC 2019，²。该方法在ISIC 2016数据集上的准确性，平均准确性，敏感性和特异性分别为85.2％，66％，52％和93.4％。获得的ISIC 2019数据库的这些评估指标分别等于96.7％，95.1％，96.3％和97.1％。此外，该方法在PH ²数据集上的准确性，敏感性和特异性分别为97.5％，100％和96.88％。根据实验结果，集成方法与每个阶段相比分别提高了评估指标。此外，与以前的研究相比，该方法可以提高黑色素瘤的检测性能。