Machine Learning for improving total burn surface area estimation

Burn injuries are a common presentation to the accident and emergency department in the UK and account for a significant cost to the NHS. The accurate assessment of these injuries by determining area and depth can ensure the patient receives the most appropriate treatment. If these assessments are incorrect, it can result in inadequate treatment or unnecessary transfer to specialists' centres causing distress to the patient and a significant cost to the NHS. The accuracy of the initial assessment can vary depending on the experience of the assessor. This study explores if machine learning methods can aid in a more accurate diagnosis of these burn injuries which may in future help to develop models that can be used in clinical practise to aid clinicians. The initial stage will assess how accurately specialists can assess burn injuries compared to a true calculated body surface area. The second stage will assess if a new model can be created to determine the difference between images of normal skin and a burn injury. This will be through a deep learning approach. The third stage will assess if a model can be created to determine the difference between full thickness burns, partial thickness burns and normal skin. Finally, we will determine if a code can be created to extract the burn from an image of burn and normal skin. Initial results have shown that specialist burn surgeons have a tendency to overestimate burns. We have also been able to develop a model that is able to accurately place a burn into the correct category 97% of the time when compared to images of normal skin.

URI

http://hdl.handle.net/10454/20204

Type

Thesis

Qualification name

MD