Benign/Malignant Skin Melanoma Detection from Dermoscopy Images using Lightweight Deep Transfer Learning

Skin is one of the largest sensory organs in the human physiology and the abnormality in skin will lead to various complications. Skin melanoma is one of the harsh conditions and the untreated melanoma will lead to death due to the cancer. Early diagnosis and treatment is essential and the clinical level detection is done using the dermoscopy. This work aims to implement the deep learning based benign/malignant melanoma classification. The various phases in this tool includes; data collection and preprocessing, deep-features mining, features reduction using the Bat Algorithm (BA), and classification and performance verification. The proposed work considers the lightweight deep-learning tool to examine the chosen image database. During this task, the MobileNet-variants and the NasNetvariants are considered for the study. The classification executed using the MobileNetV2 provided an accuracy of $92.50 \%$ and the NasNetMobile based detection offered an accuracy of $\mathbf{91.50\%}$. The serially integrated deep-features based detection helped to get $98 \%$ accuracy when the Support Vector Machine classifier is considered. This confirms that the implemented scheme provided better detection accuracy.