BurnSafe : Automatic Assistive Tool for Burn Severity Assessment by Semantic Segmentation

Abstract

Burns present a significant threat to the human life and quality of life, with high rates of morbidity and mortality. Accurate diagnosis, including the assessment of the burn area and burn depth, is essential for effective treatment and can sometimes be lifesaving. However, access to specialized medical personnel can be challenging in some cases, especially for remote or underdeveloped areas. To alleviate the burden on medical staff, researchers are exploring automatic diagnostic aid tools. Among various aspects that are corroborated in the decision on the burn diagnosis, the most important are the severity of the burn and the percentage of body surface area affected. From a computer vision point of view, one requires the semantic segmentation of any given image containing burns in order to establish the burned area and the severity class of the burn. In collaboration with medical personnel, we have a collected set of in–situ images from a local children’s hospital that were then annotated by surgeons specializing in skin burns. To compensate for the limited amount of data, a domain adaptation technique is used, namely eigen-color description for soft pseudo-labeling. The developed deep learning module is aimed to be integrated within an assistive application, called BurnSafe, that could both help and educate people and respectively reduce the workload of medical staff.