33 Automated Burn Assessment using Deep Learning and Computer Vision

Le résumé fourni par la source

Accurate diagnosis of burn size and depth guides decision-making for resuscitation and triage to burn centers. Inaccuracies in initial burn assessment by non-specialists in emergency departments may lead to inappropriate initial treatment. An automated burn assessment tool may address this problem without involving burn specialists. Deep learning and computer vision present an opportunity to provide automated size and depth assessments using digital images of burns. Digital images of 793 acute burns were collected from Google images and a regional burn center. Edges of burn images were manually segmented. The resulting 1570 individual burns were labeled for depth by three independent reviewers. The 1501 burns with majority agreement were used for training, validation, and testing of a deep learning algorithm. This algorithm performed a semantic segmentation task of whole images in which each pixel within an image was assigned a depth. This algorithm was developed for several binary classification tasks, including evaluating for superficial partial thickness or deeper burns. The performance of the algorithm was assessed using standard metrics including intersection over union (IOU), pixel accuracy (PA), receiver operating characteristic curve area under the curve (ROC-AUC), and precision-recall curve average precision (PR-AP). In 96% of the burns, a majority of independent reviewers agreed on the depth of the burn. The interrater reliability between the labelers yielded an overall Fleiss kappa of 0.60. When performing a binary class evaluation for superficial partial thickness or deeper burns, the algorithm had excellent performance across each of the commonly reported metrics including IOU 0.82, PA 0.94, ROC-AUC 0.98, and PR-AP 0.91. The included figure represents average performance of the algorithm. There were expected improvements or declines in performance with simpler or more difficult classification tasks, respectively. This study represents the largest known dataset of annotated burn images. A computer vision algorithm using deep learning achieved accurate performance on automated semantic segmentation of burns. This computer vision algorithm allows for automated identification of clinically meaningful burns of superficial partial thickness or deeper. Deploying this tool to first line providers, who could easily access this tool with a smartphone, may allow for more accurate initial management of patients with burns.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.