79. Evaluation of Multispectral Short-wave Infrared Imaging and Machine Learning Algorithms for Assessment of Burn Wound Depth

PURPOSE: Currently, the standard of care to assess burn wound depth is to rely on the subjective surgeon evaluation which has been found to be correct in only 60% of cases, resulting in 40% of burn patients being incorrectly diagnosed and receiving inappropriate surgery. Several technologies for imaging-based assessment of wound depth have been developed to examine burn wounds. However, these current modalities (Laser Doppler, NIR imaging) suffer from limitations in use and accuracy. Here we present multispectral short-wave infrared (SWIR) imaging and machine learning (ML) classification algorithms of human burn injury as a first step toward a new non-invasive and label-free technique for burn depth determination. METHODS: Prior to burn excision, a novel system, based on a specialized camera, imaged the burn areas and normal skin at 5 different SWIR narrow bands centered around 1200, 1500, 1650, 1940, and 2250 nm. Visual photographs from imaged areas were collected and presented to 5 board-certified surgeons for independent classification of burn depth (superficial, deep partial or full thickness). In the SWIR images, regions of interest (ROIs) were selected in burned and adjacent normal skin areas and the reflected light intensity in each ROI was averaged and normalized to the normal skin ROIs. The normalized intensities at the five SWIR bands for each ROI were grouped based on the consensus depth category classified by the surgeons. A preliminary MobileNet convolutional neural network (CNN) was developed to predict the consensus burn depth category from SWIR images. The CNN was trained on a randomized set of 85% (1169 images) of total samples (1377 ROIs) while the remaining 15% (208 images) was reserved for testing. To improve classification from the supervised CNN’s predictions, unsupervised ML and artificial intelligence approaches are being developed to classify the combination of SWIR and visual images based on texture-based features independent of the surgeons’ classifications. RESULTS: Multispectral images of 267 ROIs (89 full thickness, 85 deep partial thickness, 29 superficial burns and 64 normal skin) were identified from 48 burn areas in 26 patients. The total number of samples (across 5 spectral bands, for all ROIs) was 1377. Previously, we found significant differences in reflectance intensities between wound depths at 1200, 1650 and 2250 nm. Accuracy of the initial CNN model was >80% overall and 94% true positives for burns when binary classification of burn vs. normal skin was tested. This adapted model showed average class accuracies of 0.91, 0.46, and 0.18 for operable burns (deep partial/full thickness burns), normal skin, and superficial burns (superficial/superficial partial thickness) respectively. CONCLUSION: The new combination of non-invasive and label-free SWIR imaging and ML is a promising approach for accurate burn wound depth assessment. The results motivate further studies toward SWIR imaging and ML model optimization for increased surgical precision in tissue debridement.