A Majority Voting based Ensemble Approach of Deep Learning Classifiers for Automated Melanoma Detection

Abstract

Melanoma is one of the most threatening types of skin cancer with a large-scale mortality rate. The early-stage diagnosis of melanoma can help prevent its proliferation to other body organs. In this regard, several Computer aided Diagnosis (CAD) approaches have been proposed by researchers which serve as a milestone in combatting this ugly disease. In this research work, we have proposed a Model Blending technique which is an ensemble of two pre-trained deep Convolutional Neural Networks (CNNs) namely DenseNet-201 and ResNet-50. The proposed ensemble approach plays a key role in accurate melanoma classification with reduced generalization error of the two Neural Networks (NNs). Region of Interest (ROI) segmentation and lesion classification are performed using multiple, standardized dermoscopy datasets accessed from PH², Med-Node and DermIs archives. We have applied advanced data purification techniques to remove occlusions, unwanted artifacts and to adjust low contrast or illumination effects. ROI (lesion) segmentation is done using the K-means clustering algorithm which precisely separates the foreground and background pixels of the images. Affine Image Transformation and Color Space Transformation approaches are applied to augment our image datasets. The ensemble model of ResNet-50 and DenseNet-201 performed binary lesion classification (benign or malignant) using a Majority Voting technique. Our proposed segmentation method displayed satisfactory results with a precision of 91%, AUC 89%, specificity 95.7% and sensitivity score of 94%. In the classification task, the pre-trained ensemble model recorded superior results as compared to other ultra-modern melanoma diagnosis approaches and achieved an accuracy of 95.2%, specificity 96.7%, sensitivity 92.8% and AUC 98.5% on multiple dermoscopy image datasets. The evaluation results are carefully analyzed and compared with other existing melanoma detection approaches which indicate the reliability and robustness of our model.