Towards Dermatoscopic Imaging and CNN Based Accurate Skin Cancer Detection

Skin cancer is the most common and deadliest type of cancer from an incidence viewpoint. Melanomas are lethal skin cell malignancies and among the deadliest forms of cancers globally, however early diagnosis significantly enhances treatment results, thus detection poses an excellent challenge as benign lesions appear to be visually similar to malignant lesions. In order to overcome this problem, we design an automated skin cancer detection system by using CNN. The proposed model is trained and validated using the Skin Cancer MNIST: HAM10000 dataset, including 10,015 DIs, which were acquired from Kaggle for the automated diagnosis and categories of skin cancer in seven different categories of skin lesions, namely AKIEC, BCC, BKL, DF, NV, VASC, and MEL. Generalization was achieved through applied data augmentation methods. Training is done on 80% of the dataset with the validation on the remaining 20%. Our proposed CNN model has achieved the precision accuracy score of 98.88%. Thus, our proposed CNN model is above the SOTA methods, including ELM [2], VGG-16 [4], and many other existing CNN models [5], [6], [8]. Our future plan includes incorporating more data sources, along with using techniques based on TL, to upgrade the robustness of the model and its accuracy for even more effective diagnostics of skin cancer.