A detailed analysis of skin burn diagnosis through segmentation and severity visualization

Abstract

Human skin burns are the important causes of disability, illness, and death in developing countries. Moreover, in many emergencies, there were no qualified human resources for burn injury management. Determining the patient burn severity depth and delaying in decision making are the main tasks identified by the doctors and nurses when caring for any burn patients in the healthcare facility. Recently, an automatic burn assessment (emergency management) system is introduced that can determine the burn severity and improve overall survival, and quality of life through visualization. Also, the injured body surface area can be specifically indicated via touch screen interface and the combined values are estimated using Total Body Surface Area (TBSA) affected by the burn. Among these facts, the burn degree prediction system faces many problems such as inaccurate dense pose estimation, burn misclassification during training, lack of deep features, does not attain accurate results in the detection process due to the deviations in the pose, pose alignment factors, etc. Therefore, an accurate process of burn depth assessment is required to detect, segment the region of burn from the acquired images. The main goal of this thesis focused on automatic segmentation of skin burn and classifies according to the degree of severity. In order to achieve this, we introduce the Densemask Region Convolutional Neural Network (Densemask RCNN) approach for the accurate estimation of human pose and generate the human burn mask for discriminating the normal and burnt regions. This model employs the Resnet model as the backbone network for the dense feature extraction. For the deep stack connection, a weighted mapping module is utilized with different weight values that enhance the prediction accuracy. Then object detection is performed by first identifying the human pose and detecting different parts of human body parts. newline