Challenges in building image-based diagnostic support deep-learning algorithm for acute burns. (13th November 2019)

• Record Type: Journal Article
• Authors: Boissin, C; Fransén, J; Huss, F; Wallis, L; Allorto, N; Laflamme, L; Lundin, J
• Is Part Of: European journal of public health. Volume 29(2019)Supplement 4
• Publication Date: 2019-11-13
• Subjects: Epidemiology -- Europe -- Periodicals; Public health -- Europe -- Periodicals; 362.109405
• DOI: 10.1093/eurpub/ckz185.127
• Languages: English
• ISSNs: 1101-1262

Abstract

Abstract: Background: Acute burns are complex to diagnose and erroneous assessments impact on the victim's mortality and morbidity. Specialists operate from a small number of burns centres and the capacity to provide timely assistance to front line clinicians is therefore limited. Diagnostic assistance through artificial intelligence could be an option to provide a timely and equitable diagnosis. This project sheds light on the feasibility of the development of an artificial intelligence algorithm for assisted diagnosis of acute burns and clarifies challenges faced along the way. Methods: A bank of images has been built from a number of burn centres in South Africa and is continuously being updated (currently about 1200 images). Attempts have been made to train deep learning algorithms to diagnose the burn depth, an element that is challenging both at bedside and using image-based teleconsultation. We came across methodological challenges that need further consideration. Results: Some challenges are clinical, i.e. inherent to the complexity of burn diagnosis, and imply the need of having an accurate diagnosis for the burn images on which the algorithm is trained. Other challenges pertain to the actual development of an algorithm. Prior to identifying burn depth a complex task relates to the feasibility of finding the burn itself in images of varying body parts and backgrounds. Further, training an algorithm for diagnosing burn depth also requires, large numbers of varying cases, the accurate labelling of the wound area for training, and decisions to be made as regards the best outcome to train upon. Current preliminary results indicate satisfactory identification of the burn area and promising results with regards burn depth diagnosis. Conclusions: Development of artificial intelligence algorithms require strong collaborations and discussions between technical and clinical experts but are showing promising results. Key messages: Development of clinical image-based automated diagnosis involve a number of critical challenges on both the technical and clinical sides that need to be addressed prior to optimization. Once optimally developed, deep learning algorithms are a potential solution to assist with the reduction of the burden of burns on the health services by providing timely, and cost-effective advice.