115 Predicting Graft Failure in a Porcine Burn Model of Various Debridement Depths via Laser Speckle Imaging and Deep Learning

Abstract Introduction Early excision and grafting (E&G) remains a mainstay in the treatment of burns. Procedures to remove necrotic tissue from severe burn wounds continue to be challenging and may affect rates of successful grafting. The use of laser speckle imaging (LSI) may help detect necrotic tissue remaining but requires human interpretation. Additional decision-making support is needed, especially in prolonged field care settings. The purpose of this study was to evaluate whether graft success or failure can be predicted from LSI using machine learning (ML) and deep learning (DL) techniques in a porcine burn model of various burn depths. Methods Anesthetized Yorkshire pigs (n=12) were burned with a 5x5 cm square brass block (0.4 kg/cm2) @ 100°C to produce partial, deep partial, or full thickness burns (PT, DPT, FT) at 10 square sections on the back of each pig. Debridement was randomized from 1 (0.030”) to 4 (0.120”) passes performed using a dermatome, and then meshed split thickness skin grafts taken from 4 caudal donor sites were applied to wounds. Post-debridement was denoted as Day 0. Graft success/failure (>70% graft take) was determined at Day 7. Laser speckle images were captured at Days 0, 3, 7, 10, and 14. ML and DL were used to develop models in order to predict graft failure and burn/debridement depth. Model performance was measured using loss, accuracy, and confusion matrices. Results Of 120 sections corresponding to 12 pigs, 7.5% (9/120) were not burned, 41.7% (50/120) were PT, 25.8% (31/120) were DPT, and 25.0% (30/120) were FT burns. Graft failure was 19.2% (23/120), with a 50.0% (10/20) rate for DPT burns involving 1- or 2-pass debridement. Both ML and DL used 600 images for algorithm development; 540 images, for training; and 60 images, for testing. DL was superior over ML in test accuracy (93.3% vs 75.0%, p< 0.05). A three-stage architecture plus image resizing was employed in DL to predict graft failure and burn/debridement depth. A best convolutional neural network for graft failure prediction was obtained, yielding a minimum cross entropy loss of 11.7% and accuracies of 96.5% (training), 93.3% (testing 1), 100.0% (testing 2), and 96.2% (overall). Promising results were also obtained for predicting burn/debridement depth: 97.6% (training), 93.3% (testing). Conclusions This study showed that graft success or failure can be predicted from LSI and DL in a porcine burn model of various debridement depths. Use of this technology may provide a potential approach for accurately assessing early E&G in severely burned patients and may aid providers during prolonged field care scenarios. Applicability of Research to Practice Use of LSI in conjunction with DL may provide a potential approach for accurately assessing early E&G in severely burned patients and may aid providers during prolonged field care scenarios.