Real-time Classification of Skin Cancer on an Edge Device

Skin cancer represents a significant health concern, necessitating early detection as it plays a significant role in successful treatment and improved patient outcomes. This study explores the potential of deep learning technology in enhancing skin cancer detection through convolutional neural network (CNN) models. It aims to develop an optimized skin cancer detection system using three popular CNN models: Efficient-NetB4, InceptionV3, and MobileNetV2. The main objectives of this research involved evaluating the performance of these CNN models in classifying three types of skin cancer: Melanoma, Nevus, and Basel Cell Carcinoma, as well as selecting the most accurate model for deployment on a Raspberry Pi for real-time classification. The ISIC-2019 skin cancer dataset was used to train and evaluate the three CNN models. The training process encompassed comprehensive data preprocessing, image augmentation, and model fine-tuning, all implemented within a deep learning framework. Model performances were compared based on multiple metrics, including accuracy, precision, recall, and Fl-score. Among the three CNN models, MobileNetV2 demonstrated the highest accuracy of 91% in skin cancer detection, making it the most suitable choice for deployment on the resource-constrained Raspberry Pi. The conversion of the model to TensorFlow Lite format facilitated efficient utilization of computational resources, enabling real-time classification on the edge device. MobileNetV2's effectiveness as a CNN model for skin cancer detection was thus demonstrated, providing a robust solution for real-time classification on the Raspberry Pi. The deployed system becomes well-suited for on-site and remote screening applications, contributing to the field of skin cancer detection by combining accurate classification and portability.