Input space augmentation for skin lesion segmentation in dermoscopic images

Cancer is the second leading cause of death globally, and of all the cancers, skin cancer is the most prevalent. Early diagnosis of skin cancer is a crucial step for maximizing patient survival rates and treatment outcomes. Skin conditions are often diagnosed by dermatologists based on the visual properties of the affected regions, motivating the utility of automated algorithms to assist dermatologists and offer viable, low-cost, and quick results to assist dermatological diagnoses. Over the last decade, machine learning, and more recently, deep learning-based diagnoses of skin lesions have started approaching human performance levels. This thesis studies approaches to improve the segmentation of skin lesions in dermoscopic images, which is often the first and the most important task in the diagnosis of dermatological conditions. In particular, we present two methods to improve deep learning-based segmentation of skin lesions by augmenting the input space of convolutional neural network models. In the first contribution, we address the problem of the paucity of annotated data by learning to synthesize artificial skin lesion images conditioned on input segmentation masks. We then use these synthetic image mask pairs to augment our original segmentation training datasets. In our second contribution, we leverage certain color channels and skin imaging- and illumination-based knowledge in a deep learning framework to augment the input space of the segmentation models. We evaluate the two contributions on five dermoscopic image datasets: the ISIC Skin Lesion Segmentation Challenge 2016, 2017, and 2018 datasets, the DermoFit Image Library, and the PH2 Database, and observe performance improvements across all datasets.