Improved Melanoma Recognition using Score Fusion Framework based on Deep Classifiers

Abstract

Melanoma is a fast-growing and malignant cancer that affects neural crest-derived cells of the skin. Early detection is the key to survival and rapid recovery; however, typical diagnosis is based on visual inspection by expert dermatologists. Many melanoma recognition methods have been proposed in the literature that can classify lesions, based on hand-crafted as well as deep learning-based features. However, an accurate, automated method for melanoma is still required. This research is aimed at improving the accuracy of melanoma recognition by deploying an ensemble of deep classifiers. Three different deep CNNs (AlexNet, ResNet, and GoogLeNet) were used for classification after fine-tuning or feature extraction. The scores of individual classifiers were then aggregated using a score fusion method. The performance of the method has been measured on the PH-2 dataset with an accuracy (ACC), sensitivity (SN), specificity (SP), dice similarity coefficient (DSC), and Area under ROC (AUC) of 98.33, 91.97, 100, 95.65 and 98.96% respectively.