Department of Electrical and Computer Engineering, Biomedical Engineering Group, Tarbiat Modares University, Tehran, Iran.
Skin Res Technol. 2013 Feb;19(1):e113-22. doi: 10.1111/j.1600-0846.2012.00617.x. Epub 2012 Jun 5.
BACKGROUND/PURPOSE: Melanoma is the most dangerous type of skin cancer, and early detection of suspicious lesions can decrease the mortality rate of this cancer. In this article, we present a multi-classifier system for improving the diagnostic accuracy of melanoma and dysplastic lesions based on the decision template combination rule.
First, the lesion is differentiated from the surrounding healthy skin in an image. Next, shape, colour and texture features are extracted from the lesion image. Different subsets of these features are fed to three different classifiers: k-nearest neighbour (k-NN), support vector machine (SVM) and linear discriminant analysis (LDA). The decision template method is used to combine the outputs of these classifiers.
The proposed method has been evaluated on a set of 436 dermatoscopic images of benign, dysplastic and melanoma lesions. The final classifier ensemble delivers a total classification accuracy of 80.46%, with 67.73% of dysplastic lesions correctly classified and 83.53% of melanoma lesions correctly classified.
The results show that the proposed method significantly increases the diagnostic accuracy of dysplastic and melanoma lesions compared with a single classifier. The total classification rate is also improved.
背景/目的：黑色素瘤是最危险的皮肤癌类型，早期发现可疑病变可以降低这种癌症的死亡率。在本文中，我们提出了一种基于决策模板组合规则的多分类器系统，以提高黑色素瘤和发育不良病变的诊断准确性。
首先，将图像中的病变与周围健康皮肤区分开来。接下来，从病变图像中提取形状、颜色和纹理特征。将这些特征的不同子集输入到三个不同的分类器中：k-最近邻（k-NN）、支持向量机（SVM）和线性判别分析（LDA）。使用决策模板方法组合这些分类器的输出。
该方法已在一组 436 张良性、发育不良和黑色素瘤病变的皮肤镜图像上进行了评估。最终的分类器集成提供了 80.46%的总分类准确率，其中 67.73%的发育不良病变被正确分类，83.53%的黑色素瘤病变被正确分类。
结果表明，与单个分类器相比，所提出的方法显著提高了发育不良和黑色素瘤病变的诊断准确性。总分类率也得到了提高。
