Real-Time Recognition of Malignant Skin Lesions using Ensemble Modeling

Similarities between early malignant skin lesions and benign skin lesions make it challenging to correctly differentiate between the two. In this paper, a hybrid method for classification of skin lesions using an ensemble of deep predictive models (Feedforward Neural Network, Image Histogram Classification using KNN, LeNet-5 CNN architecture, VGG-11 CNN architecture) is proposed to correctly identify Malignant skin lesions. The image dataset contains multiple images of Benign (Nevus, Seborrheic Keratosis) and Malignant (Melanoma, Carcinoma) skin lesions supplied by the ISIC (International Skin Imaging Collaboration) Archive, an open source organization aiming for progressive studies in skin cancer detection. The challenge is to correctly segment the region of interest (skin lesion) in the initial stages, despite hairy images. We develop an ensemble of three models which are not correlated with each other (a pre-requisite for ensemble modeling). We also develop a MATLAB (64-bit R2016a, The Math Works) based Graphical User Interface (GUI) which provides real-time results for uploaded skin lesion images. The aim of the paper is to propose a method for early detection of melanoma, which is not far behind the time-consuming traditional biopsy procedures.