A Transformer-Based Computational Framework for Automated Burn Injury Segmentation and Severity Assessment

Burn injuries are the fourth most common type of injury; they are usually caused when the skin comes in contact with fire, hot water, radiation, or dangerous chemicals. According to the World Health Organization (WHO), an estimated 180,000 deaths occur every year due to burns. Different types of burn injuries require different medical procedures and medications; therefore, it is essential to diagnose a burn injury as early as possible to ensure that the patient receives proper treatment. Burns are usually diagnosed through visual inspection, which is subjective to the doctor's experience and is usually only 50-70% accurate. A delay in diagnosis and precise treatment may lead to severe complications and, in worst-case scenarios, may also cause death. Unfortunately, in remote areas of Pakistan, where only basic dispensaries, clinics, or Primary Healthcare Units (PHUs) are available, and due to the lack of burn experts and the absence of Laser Doppler Imaging (LDI) equipment, such accurate diagnoses are not possible. As a result, many burn victims are either left untreated or given incorrect treatment, leading to complications. In this research, we propose a deep learning-based approach to assess the severity of acute burns in terms of burn degree and the affected Total Body Surface Area (TBSA%) from 2D images. The primary objective of our work is to determine the degree of a burn injury (first-degree, second-degree, third-degree, fourth-degree, and mixed-degree) and, based on the classification, estimate the TBSA% to advise on the appropriate medication required for the affected area. To achieve real-time diagnosis, an Android-based system has also been developed, allowing users to capture images and obtain instant results on acute burns detected in the image, including their severity, TBSA%, and recommended treatment.