Malignant skin melanoma detection using image augmentation by oversampling in nonlinear lower-dimensional embedding manifold

The continuous rise in skin cancer cases, especially in malignant melanoma, has resulted in a high mortality rate of the affected patients due to late detection. Some challenges affecting the success of skin cancer detection include small datasets or data scarcity problem, noisy data, imbalanced data, inconsistency in image sizes and resolutions, unavailability of data, reliability of labeled data (ground truth), and imbalance of skin cancer datasets. This study presents a novel data augmentation technique based on covariant Synthetic Minority Oversampling Technique (SMOTE) to address the data scarcity and class imbalance problem. We propose an improved data augmentation model for effective detection of melanoma skin cancer. Our method is based on data oversampling in a nonlinear lower-dimensional embedding manifold for creating synthetic melanoma images. The proposed data augmentation technique is used to generate a new skin melanoma dataset using dermoscopic images from the publicly available P H2 dataset. The augmented images were used to train the SqueezeNet deep learning model. The experimental results in binary classification scenario show a significant improvement in detection of melanoma with respect to accuracy (92.18%), sensitivity (80.77%), specificity (95.1%), and F1-score (80.84%). We also improved the multiclass classification results in melanoma detection to 89.2% (sensitivity), 96.2% (specificity) for atypical nevus detection, 65.4% (sensitivity), 72.2% (specificity), and for common nevus detection 66% (sensitivity), 77.2% (specificity). The proposed classification framework outperforms some of the state-of-the-art methods in detecting skin melanoma.