Applying deep learning to detect melanoma using mobile phone cameras

Abstract

Malignant melanoma, a severe form of skin cancer has seen an increase in incidence, emphasising the need for early detection and diagnosis. This thesis investigates the potential of convolutional neural networks (CNNs) in classifying moles as malignant or benign using images captured from mobile phone cameras. The primary goals were to assess the performance of these models and to develop a mobile application that utilizes the trained models to provide preliminary assessement of potentially cancerous moles. The ISIC archive was used as a dataset, with over 75 000 images of moles labeled either malignant or benign, to train and validate the models. Through extensive hyperparameter optimization using the Optuna framework, in addition to fine tuning techniques, the EfficientNetV2B3 model achieved a validation accuracy of 86.73%. A mobile application was developed to integrate the trained model, allowing users to capture images of moles and recieve real time classification results. The application’s usability varied due to mixed conditions when taking images, highlighting the importance of optimal conditions. The results indicate that in the right conditions, the mobile application might be able to effectively classify images, serving as a useful tool for early melanoma detection. This thesis concludes that deep learning models, combined with mobile technology, hold significant promise in improving skin cancer diagnosis in the future.