Aline Okita
Most articles evaluate algorithms accuracy to classify pigmented lesions. However, highest accuracy are not always related to the lowest health system costs. One melanoma false negative (FN) can lead to later stages disease resulting in highdirect and indirect costs. We worked with a dataset of 8000 dermoscopy images from ISIC dataset (4000 melanomas). 80% of them were used for training, 10% for validation and 10% for testing. We transferred learning from a pre-trained image net to Efficient Net and trained it with 20 epoch. We built a cost matrix based on values spent by the Health Public System in Brazil for the diagnosis and treatment of melanoma in different stages and measured how the algorithm would perform in a real world set of patients according to the cost of each wrong prediction.