AI-Based Segmentation of Wound Types Caused by Epidermolysis Bullosa

Epidermolysis bullosa (EB) defines a group of rare, inherited and currently incurable genetic disorders characterized by excessive skin fragility, with blistering and wounding of skin and mucous membranes upon minor mechanical trauma. Accurate wound assessment is imperative for measuring and longitudinal monitoring of disease activity and for determining the most accurate treatment. In this work, we describe the annotation of clinical images from EB patients and the subsequent training of both classical convolutional neural networks and state-of-the-art transformer-based architectures for the segmentation of various categories of EB skin wounds. Utilising our dataset of 260 EB images from 18 patients and the corresponding 536 expert-annotated segmentation masks, we train and evaluate five different model architectures and compare their performance against human inter-annotator agreement. External evaluation was not possible because no comparable annotated EB datasets are available. Our results demonstrate that transformer-based models, particularly Mask2Former, achieve near-expert-level segmentation accuracy in five of the seven categories and even surpass human inter-annotator agreement in two. Models based on convolutional neural networks perform noticeably worse and generally fail to accurately segment rarely occurring classes. Averaged over all categories, Mask2Former achieved a mean dice similarity coefficient (DSC) of 52.9%, compared to an inter-annotator agreement of 55.9%. These results indicate that Mask2Former has the potential to serve as a clinical decision support system to improve wound segmentation and categorization. Furthermore, this model provides a foundation for planned future applications in automated wound measurement, wound progression monitoring and documentation and the development of a smartphone-based telemedicine application.