SkinSight: A Melanoma Detection App Based on Deep Learning Models with On-Device Inference

A timely and accurate skin cancer diagnosis is a key factor in reducing mortality rates, especially with melanoma which often resembles in its early stages with moles. Convolutional neural networks (CNNs) are models commonly used to classify dermoscopy images into benign or malignant. CNNs are frequently implemented on Graphical Processing Units (GPUs), which are not always available in rural areas. This paper compares three CNNs to classify benign and malignant melanoma images. We select the most appropriate neural architecture by comparing accuracy results and model lightness to load it on a mobile device. With this strategy, the training of the CNN is performed on the GPU and the inference in portable devices that can be used in rural areas. The developed app is named SkinSight. This app was evaluated with images of two different datasets achieving competitive results compared to state-of-the-art models. Considering that most people have a mobile device, this app could be used in areas where it is difficult to have specialized GPUs and highly trained personnel in cancer detection.