A deep learning based patient care application for skin cancer detection

Melanoma is one of the deadliest skin cancers. It can, however, be cured with a high success rate if found in the early stage. With the recent development in Artificial Intelligence (AI), there has been an increase in the research of Deep Learning (DL) models for melanoma detection. However, such deployments are primarily trial-based and still to gain wider acceptance from the practitioners. One of the primary reasons for such lower acceptance lies in the inherent approach and ethical concerns. To this end, our study designs a Deep Learning-based recommendation application for patients to advise whether they need to see a dermatologist based on the provided skin lesion image. The traditional AI-based melanoma classifiers tend to be developed based on bulk image training and hence would classify images on an individual level. However, they neglect that a single skin lesion is merely a part of the patient. This research considers the contextual information between different lesions on the same patient. It produces a personalized patient classification. This is achieved by making preliminary image-level predictions and then fusing them into holistic patient-level classification. Models are developed using transfer learning and deep neural networks. These models are then tested on a real melanoma dataset with the true label provided. An external benchmark was also established from similar recent studies using the same dataset. Using the patient's unique biological features, the best classifier achieves an 81% Area Under the Receiver Operating Characteristic Curve (AUROC) score.