Depthwise Separable Convolutional Neural Network for Skin Lesion Classification

Melanoma is one of the deadliest skin cancers. Early diagnosis plays an essential role in effective treatment planning and reducing the mortality rate of skin cancer. In this study, we propose a compact deep learning-based classification model with a separable convolutional neural network for melanoma detection. The proposed architecture is aimed to minimize the need for task-specific data pre-processing methods such as noise reduction, artifact removal, and low contrast adjustment to support a better generalization ability for unseen / test data. We validated the performance of the proposed method on the ISIC 2018 dataset. Our results show that the proposed architecture achieves comparable accuracy to the widely-used architectures presented in the literature while being more compact and simpler. The proposed methodology achieves 87.24% accuracy, 95.94% sensitivity and 98.47% specificity.