Enhancing burn wound classification with ensemble learning using spatial frequency domain imaging applied to in-vivo porcine skin

Assessing burn severity is crucial for appropriate care and treatment to reduce the risk of scarring and infection. Currently, practitioners rely on visual and tactile assessments, which can be subjective and potentially misleading. A more objective method recently employed in a swine model involves spatial frequency domain imaging (SFDI), a wide-field, noninvasive technique that deduces tissue optical properties using light propagation models. Recent studies have advanced this approach by using machine learning (ML) algorithms to predict burn severity from calibrated reflectance data obtained via SFDI, eliminating the need for light propagation models. Data from a porcine model of graded burns was collected using two distinct SFDI systems: the Reflect RS and Clarifi RS (Modulim Inc.). The Reflect RS operates with eight wavelengths from 470 to 850nm modulated at six spatial frequencies, five from 0 to 0.2/mm at equal increments and another at 0.122/mm, while Clarifi RS uses five wavelengths in the same range with modulation only at 850nm (0 and 0.122/mm). This data was previously used exclusively to train simple ML models. In this study, we evaluate and compare ensemble learning algorithms against a support vector machine (SVM) for classifying burn severity in pigs. Specifically, we used random forest and Libra, the latter commonly used as a low spectral resolution brain tumor classifier for medical hyperspectral imaging. To ensure fair comparison, our methodology avoids overfitting, differing from previous studies. Our results demonstrate that EL-based models can improve performance by up to 20% over SVM in metrics like F1-Score. Additionally, EL models generate more accurate classification maps with fewer false positives, identifying areas likely to heal after 28 days, thus offering clear and practical clinical insights.