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

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