Robust Visual Identification of Under-resourced Dermatological Diagnoses with Classifier-Steered Background Masking.

Collecting images of rare dermatological diseases for machine learning detection applications is a costly, laborious task. It is difficult to collect enough images of these diagnoses to avoid the risk of low accuracy "in the wild". One of the sources of bias in these networks is irrelevant background pixel data. These pixels necessarily have no clinical significance, yet Deep Neural Networks will make weak correlations based on that information. To reduce their ability to do this, we introduce a masking augmentation algorithm, InfoMax-Cutout. It employs unsupervised Information Maximization losses to mask out background pixels. InfoMax-Cutout increased accuracy on classifying 319 diagnoses by 0.76%. These features generalized to an unseen diagnosis task (Fitzpatrick 17k), improving accuracy over a baseline by 43.3% and reducing Gini inequality by 20.9%. This approach of learning to separate out background pixels can increase accuracy in detecting diseases in Lower and Middle Income Countries.