Deep learning in clinical dermatology

The prevalence of skin diseases is high. A recent survey reported that half of the European population was afflicted with skin conditions. However, the resulting demand for dermatological care cannot be met satisfactorily because of a general shortage of dermatologists that will realistically not be filled by the healthcare sector. Alternative solutions should therefore be pursued to increase the capacities of the current healthcare workforce.
The recent progress of machine vision enabled by deep learning has allowed researchers to automate parts of dermatologists' workflow with an effective scale-up potential. In this work, we present different approaches based on deep learning that either include aspects of dermatologists' workflow or whose predictions can easily be verified by clinicians. We propose a method for the generation of anatomical maps from patient photographs to assist dermatologists with lesion documentation and enable lesion detection and segmentation systems to stratify their predictions anatomically. Based on key features from lesion dermatological description, we develop an approach for the differential diagnosis of skin diseases.
To enable objective severity assessment, we propose a method for the segmentation and quantification of palmoplantar pustular psoriasis, ichthyosis with confetti and hand eczema. Combined with the anatomy approach, we generate the anatomical stratification of hand eczema lesions. To concretize our research efforts, we present an African teledermatology initiative aiming to provide semi-automatic triage of the six most prevalent local skin diseases. Finally, we introduce our framework to enable researchers with medical background to train and evaluate deep learning models.