521 Non-Invasive Imaging Device for Detection of Non-Healing Burns: An Initial Clinical Case Series

Abstract

Accuracy of burn depth assessment depends on expertise and can be as low as 64%, even for skilled practitioners. Imaging devices to classify burn depth, including laser Doppler imaging, multispectral imaging (MSI), and thermography, have been studied to address this issue. The objective of this study was to determine if a MSI device, previously developed in an animal burn model, could translate to clinical burns. We present current results of this ongoing proof-of-concept (POC) clinical study, including study design and initial burn detection accuracy. In an IRB-approved study, data were collected from subjects with 1 , superficial 2, deep 2 , and 3 thermal burns. Subjects were imaged following consent and daily for up to 7 days post injury. At imaging timepoints, the MSI device was used to collect images across the visible and near-IR spectrum. True severity of burn injuries, or ground truth, was determined using 21 day healing assessments or pathology for burns that required excision. Using MSI data and ground truth, we trained two deep learning algorithms to identify pixels in the image that represened non-healing burn tissue. A a fully connected convolutional neural network (CNN) and a fully convolutional neural network (SegNet) were trained to segment non-healing burn pixels from other pixels in the MSI images. Accuracy, sensitivity, and specificity of these algorithms ability to identify non-healing burns on the current study subjects were calculated using cross-validation (CV). Average accuracy was 75 ± 0.3% (fig. 1), and this accuracy increased as we collect more images. Additionally, classified output images could be processed in less than one second using the SegNet algorithm. Results from two deep learning algorithms on initial POC study subjects obtained using leave-one-out CV. Obtaining MSI images with the described clinical study design was feasible. The MSI images contained sufficient information to classify non-healing burn tissue as accurately as a skilled practitioner, and could provide these results rapidly. These preliminary results are promising, and future work will be aimed toward collecting more data to identify sources of variability and to increase algorithm accuracy. This study shows results from an ongoing POC study for developing an MSI device to aid in burn depth assessments. Eventually, the device could assist in EDs that lack specialized burn care and as an aid to burn surgery.