Deep learning for skin lesion segmentation

Skin cancer is a major public health problem requiring computer aided diagnosis to reduce the burden of disease's high incidence ratio and the associated expenses by assisting clinicians. Image segmentation, the task of decomposing an image into multiple regions by per pixel labeling, is a crucial step toward skin cancer diagnosis and treatments. However, the existence of natural and artificial artifacts (e.g. hair and air bubbles), intrinsic factors (e.g. lesion shape and contrast), and variation in image conditions originating from imaging tools and environments make skin lesion segmentation a challenging task. Recently, several efforts have been made to leverage the demonstrated superior performance of deep learning models in the segmentation of skin lesions from the surrounding healthy skin. In this thesis, after a thorough examination of the studies leveraging the capability of deep learning models in skin lesion segmentation, we propose novel segmentation prediction models advancing state-of-the-art skin lesion segmentation techniques. First, we introduce deep learning based models that leverage the auxiliary information in the form of domain knowledge, contextual information, and labels consistency to regularize model parameters toward a more generalizable solution. Specifically, we encode high order shape prior knowledge into the loss function and also incorporate high-level semantic information in learning a sequence of deep models. Second, we study the limitations of ground truth pixels level annotations to effectively leverage limited reliable annotations. Specifically, we propose a robust to noise network by learning spatially adaptive weight maps associated with training images encoding the level of annotation noise to reduce the requirement of careful labeling. Also, we avoid single annotator bias, by training in an ensemble paradigm that handles inter-annotator disagreements and learns from all available annotations.