Privacy-preserving With Sonification For Training of Convolutional Deep Neural Networks for Melanoma Diagnosis

Melanoma is a form of cancer that is a primary cause of skin cancer deaths. A major predictive factor for positive patient outcomes is diagnosis of disease in earlier cancer stages before the disease has spread beyond the initial lesion. However, many patients are diagnosed late because they cannot afford to meet a doctor or are embarrassed to be examined. These patients suffer from a significantly greater rate of mortality. As a remedy, machine learning models have been proposed to enable easy and automated diagnosis using images. However, the development of models for use in a clinical setting is not yet possible due to the limited availability of training data. Training data that is available is often private and thus isolated within individual institutions. Therefore, a large data set containing patients of different ancestries, skin colors, and ages is not available. In this study, we show that the Sonification of images results in a greater proportion of patients' consent to share their data in a public database, and that models trained from Sonified images have similar performance to those trained on raw skin lesion images.