Incorporating Diagnostic Prior with Segmentation: A Deep Learning Pipeline for the Automatic Classification of Autoimmune Bullous Skin Diseases

The term AutoImmune Bullous Diseases (AIBDs) refers to a wide group of skin disorders, in which autoantibodies are developed and directed against proteins of the epidermis and the basal membrane. The correct diagnosis and classification of AIBDs require the analysis of ImmunoFluorescence (IF) skin images. Up to now, it can only be performed by operators with great clinical expertise and it is limited to highly specialized centers. In this paper, we exploit deep convolutional neural networks and a transfer-learning framework to design an automatic classification framework for AIBDs diagnosis. We incorporate a diagnosis prior from the experts of the field (the visual analysis of the basal membrane) in an initial segmentation step that allows our network to focus on the most informative parts of the image. The designed model consists of a two-branches CNN. We report an experimental analysis based on AIBDs acquired in our lab, reaching very accurate results. Our classification pipeline, with the incorporation of medical diagnosis prior, could lead to an increase of medical trust in deep learning algorithms for the automatic classification of AIBDs skin disorders.