An Automated Deep Learning based Ensemble Approach for Malignant Melanoma Detection using Dermoscopy Images

Melanoma is categorized as an extremely lethal type of skin cancer. Its earlier and timely diagnosis is the only solution to minimize the fatality rate in patients. Various computer aided diagnosis (CAD) systems have been designed which show great advancement in lesion segmentation and classification. The Model Blending approach is the ensemble of multiple Convolutional Neural networks (CNNs) which results into lower variance in their output predictions, thus reducing the generalization error by many folds. This proposed study is designed to provide a fully automated deep learning based melanoma detection framework using multiple, standard skin lesion databases including PH2, Med-Node and ISIC-2020. Extensive pre-processing on dermoscopic images is performed to remove useless artefacts and preserve illumination effects. Semantic segmentation is carried out using a Fully Convolutional Network (FCN-8), followed by image augmentation methods. An ensemble of deep ResNet-50 and Inception-V3 has been designed to perform binary classification (benign or melanoma) of lesion images. The segmentation approach exhibited satisfactory performance with an accuracy score of 94%, Dice coefficient 88% and Jaccard similarity coefficient 89%. In the classification task, the pre-trained CNN model successfully recorded an average accuracy of 93.4%, specificity 96.5%, ROC-AUC 98.8% and average precision 89.5% on augmented dermoscopy images. The classification results of the model are deeply analyzed and compared with other ultra-modern melanoma diagnosis frameworks which indicate that our proposed model successfully achieved better segmentation and classification results. This ensemble approach is fully practicable and can be deployed by dermatologists as their medical assistant/guide.