Skin Cancer Dermoscopy Images Classification with Meta Data via Deep Learning Ensemble

The incidence of both non-melanoma and melanoma skin cancers has been increasing over the past decades. Currently, between 2 and 3 million non-melanoma skin cancers and 132,000 melanoma skin cancers occur globally each year. In this study, we proposed an ensemble CNNs for multi-class classification study based on integration of image preprocess, deep learning and risk management on skin cancer dermoscopy images with Meta-data. We basically demonstrate general stacking ensemble flow by stages from data preprocessing, CNNs grouping, Meta-data concatenating and 1st CNNs ensemble and 2nd meta-classifiers ensemble. Our result showed good performance, especially the contribution from 1st ensemble while we check break-down details. With our 5% holdout validation dataset, we saw accuracy maximum 91% in our ensemble model, compared with best single CNN (DenseNet121) 87.3%. Furthermore, we also introduced cost adjustment analysis in ensemble phase to manipulate different risk management trade-off strategy for specific categories or metrics, rather than in deep learning phase, so that no need to retrain CNN via loss function adjustment.