Classifying melanoma images with ensembles of deep convolutional neural networks

Summary

Malignant melanoma is the deadliest form of skin cancer and is one of the most rapidly increasing cancers in the world. Proper diagnosis of melanoma at an earlier stage is crucial for a high rate of complete cure. Both patient and physician awareness regarding the signs and symptoms of early melanoma remains paramount. Hence, a reliable automatic melanoma screening system would provide a great help for clinicians to detect the malignant skin lesions as early as possible. In the last years, the efficiency of deep learning-based methods increased dramatically and their performances seem to outperform conventional image processing methods in classification tasks. In this master thesis, the EfficientNet family of convolutional neural networks is utilized and extended for identifying malignant melanoma on a dataset of 58,457 dermoscopic images of pigmented skin lesions. A comparative study of the effects of different training configurations is conducted to reveal what contributes to improve performance, and all trained networks are aggregated with an ensembling strategy to further improve individual results. The proposed method has been evaluated on the SIIM-ISIC Melanoma Classification 2020 dataset and the best ensemble model achieved 0.9404 area under the ROC curve score on hold out test data.