Automatic Diagnosis of Melanoma Using Log-Linearized Gaussian Mixture Network

Abstract

Melanoma is the most malignant type of pigmented skin lesions whose early diagnosis is the only treatment key. This paper presents a decision support system for automatic melanoma recognition using log-linearized Gaussian mixture neural network (LLGMNN). Here, some image preprocessing steps precede segmentation to remove artifacts. Next Otsu thresholding method is utilized to detect lesion from the surrounding healthy skin. Then related features including shape and border characteristics, color, and texture features are extracted. A mutual information based feature selection technique is used to find the optimal subset of attributes. Here, two different structures of LLGMNN are designed and validated for our pattern classification problem, one for detection of melanoma from non-melanoma lesions and the other one for discrimination between melanoma, dysplastic, and benign lesions. The proposed system is evaluated on a set of 792 dermoscopy images. Classification results show the accuracy of 89.8%, 88.3%, and 91.2 % for melanoma, dysplastic, and benign lesions, respectively. Results show that the proposed system is efficient, and achieve acceptable classification accuracies.