Intrapatient Diagnosis for Melanoma Skin Cancer

Melanoma is the most fatal form of skin cancer, making its research a hot topic worldwide. Despite the promising results achieved by recent methods, they often disregard one of the most important criteria considered by dermatologists: the intrapatient lesion context. In this work we address this limitation, leveraging the medical concept of the ugly duckling (UD) lesion, and translating it to both the deep learning and the anomaly detection fields. Considering this, we propose an integrated diagnosis at the image and patient levels. Our results strongly support the hypothesis that the intrapatient context can enhance the diagnosis. The detection of melanoma at the image level improved by 7%, while the patient diagnosis improved by over 10% when compared with models that did not incorporate the patient context.