Feature learning and transfer learning approaches for classification of human burn wounds using multispectral SWIR imaging

Accurate determination of burn wound depth is crucial in the resection of non-viable tissue to avoid infection, complications in healing, and the unnecessary removal of healthy tissue. To improve the accuracy in burn wound depth assessment, we evaluated skin burns using novel multispectral short-wave infrared (SWIR) imaging. This technology has shown promise in determining burn wound depth by reflecting levels of skin moisture, collagen, and necrosis which can indicate tissue vitality. Multispectral SWIR images were obtained at five narrow wavelength bands between 1200- 2250 nm for 267 regions of interest (ROIs) in 48 burn areas of 27 consecutively admitted patients. 85 full thickness burns, 71 deep partial thickness burns, 28 superficial thickness burns, and 61 normal skin ROIs were classified by blind surgeons with consensus (≥ 60% agreement). A random forest (RF) classifier trained on reflectance intensity features and texture properties of gray level co-occurrence matrices showed test accuracies of 62.4% when distinguishing between non-operational and operable ROIs, and an average test accuracy of 70.2% across all classes when classifying between normal skin, superficial partial thickness burns, and operable burns. A VGG-16 feature extractor with a RF classifier and a fine-tuned VGG-16 model with fully connected layers resulted in test accuracies of 52.9% and 60.0% for binary classification, and 60.0% and 67.1% for 3-category classification, respectively. With additional data sources and the use of more objective standards for accuracy evaluation, these classification pipelines may be adapted for tools to be used by burn surgeons, emergency responders, and clinicians to support more accurate decisions for burn wound care.