Unsupervised SoftOtsuNet Augmentation for Clinical Dermatology Image Classifiers.

Abstract

Data Augmentation is a crucial tool in the Machine Learning (ML) toolbox because it can extract novel, useful training images from an existing dataset, thereby improving accuracy and reducing overfitting in a Deep Neural Network (DNNs). However, clinical dermatology images often contain irrelevant background information,such as furniture and objects in the frame. DNNs make use of that information when optimizing the loss function. Data augmentation methods that preserve this information risk creating biases in the DNN's understanding (for example, that objects in a particular doctor's office are a clue that the patient has cutaneous T-cell lymphoma). Creating a supervised foreground/background segmentation algorithm for clinical dermatology images that removes this irrelevant information would be prohibitively expensive due to labeling costs. To that end, we propose a novel unsupervised DNN that dynamically masks out image information based on a combination of a differentiable adaptation of Otsu's Method and CutOut augmentation. SoftOtsuNet augmentation outperforms all other evaluated augmentation methods on the Fitzpatrick17k dataset (0.75% improvement), Diverse Dermatology Images dataset (1.76% improvement), and our proprietary dataset (0.92% improvement). SoftOtsuNet is only required at training time, meaning inference costs are unchanged from the baseline. This further suggests that even large data-driven models can still benefit from human-engineered unsupervised loss functions.