Automated deep learning approach for classification of malignant melanoma and benign skin lesions

Abstract Skin cancer becomes a significant health problem worldwide with an increasing incidence over the past decades. Due to fine-grained differences in appearance of skin lesions, it is very challenging develop automated system for benign-malignant classification through images. This paper proposes novel Computer Aided Diagnosis (CAD) lesion high performance using accuracy low computational complexity. A pre-processing step based on morphological filtering employed hair removal and artifacts removal. lesions are segmented automatically Grab-cut minimal human interaction HSV color space. Image processing techniques investigated automatic implementation ABCD (asymmetry, border irregularity, dermoscopic patterns) rule separate malignant melanoma from benign lesions. To classify into or malignant, different pretrained convolutional neural networks (CNNs), including VGG-16, ResNet50, ResNetX, InceptionV3, MobileNet examined. The average 5-fold cross validation results show that ResNet50 architecture combined Support Vector Machine (SVM) achieve best performance. also effectiveness data augmentation both training testing achieving better than obtaining new proposed diagnosis framework applied real clinical experimental reveal superior other recent terms area under ROC curve 99.52%, 99.87%, sensitivity 98.87%, precision 98.77%, F1 -score 97.83%, consumed time 3.2 s. reveals can be utilized help medical practitioners classifying