Real-time AI Image Classification System for Burn Injuries

Abstract

With the development of Taiwan's economy and public health education, burn injuries are no longer a leading cause of death in Taiwan. Nevertheless, burn injuries are still unavoidable in our life. Thus, we must not underestimate the possibility of harmful consequences. In the past, large-scale deep learning models were employed to quickly classify the severity of burn injuries. Despite this, the recognition process still required individuals to make a self-assessment regarding the need for medical attention. However, for users without professional medical knowledge, performing these tasks can be challenging. Additionally, given the widespread use of the Internet, as most users are accustomed to accessing information via mobile devices, expecting injured users to operate a desktop computer for wound classification is unrealistic. Therefore, this study aims to develop a web-based platform for classifying burn injuries that aligns with users' current habits. Users will be able to import wound images, and the system will automatically provide information regarding the severity of the burn injury, along with relevant medical advice. As for the classification system, we selected the GTM platform to train our deep learning model for optimal integration with the web server. Finally, we have rebuilt the training image dataset, which now includes 1,162 images of burned skin and 305 images of normal skin. The precision rates can achieve over 81% to 95% for all four degrees of burn injuries during the later stages of model verification. The customer satisfaction with the classification platform also surpasses a “satisfactory” level.