Evaluating research-grade and commercial SFDI platforms for burn severity assessment and feature reduction
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Spatial frequency domain imaging (SFDI) has demonstrated the ability to provide early, quantitative assessment of burn wound severity. Previous studies using a research-grade SFDI platform showed that high-dimensional datasets incorporating multiple spatial frequencies and wavelengths can predict healing outcomes in controlled porcine models of graded burns. To assess the impact of reduced measurement dimensionality on diagnostic performance, we compared classifications derived from a research-grade SFDI system (Reflect RS) with those obtained using a simplified commercial SFDI platform (Clarifi RS). Both systems were used to image graded burns 24 hours after injury. Pixel-level classifiers were trained using regions defined by 28-day healing outcomes, and models based on the full Reflect dataset were compared with classifiers generated from reduced Reflect feature sets and datasets designed to mimic Clarifi acquisition features. Performance was evaluated using leave-one-subject-out validation. The full Reflect dataset achieved the highest classification performance, with mean F1 scores approaching 0.88. Although reducing the number of measured features decreased classification accuracy, simplified models and Clarifi-based datasets maintained F1 scores greater than 0.8 for binary classification. These findings indicate that dimensionality reduction produces a measurable but manageable loss in performance and support the development of clinically practical, application-specific SFDI systems for burn assessment.