Burn Severity Classification Using Convolutional Neural Networks With Thermal Imaging and Diffuse Reflectance Spectroscopy

This is the authors' abstract. We don't add key points for this paper.

Accurate assessment of burn severity is essential for clinical decision-making and forensic investigations. However, conventional visual examination is subjective and often cannot reliably distinguish between different burn depths. Although machine learning (ML) and deep learning (DL) techniques have improved automated burn classification, most existing studies rely on a single sensing modality and lack synchronized thermal and optical measurements acquired under identical experimental conditions. This study presents a controlled experimental framework for burn severity classification using thermal imaging and Diffuse Reflectance Spectroscopy (DRS). A controlled experimental framework was established using ex vivo porcine skin, where burn severity was systematically induced by varying the exposure time while maintaining a constant temperature. A synchronized dual-modality dataset comprising approximately 4,060 thermal images and 1,368 diffuse reflectance spectra was acquired under identical burn conditions, representing five burn conditions, including unburned tissue and four burn severity levels. Two independent deep learning models were developed and evaluated: a convolutional neural network (CNN) for thermal imaging and a one-dimensional convolutional neural network (1D-CNN) for Diffuse Reflectance Spectroscopy. The thermal imaging model achieved an overall classification accuracy of 97.17%, with precision, recall, and F1-score exceeding 97%, whereas the Diffuse Reflectance Spectroscopy model achieved an average classification accuracy of 92.91% ± ~2.93 % using five-fold stratified cross-validation. The results demonstrate that both sensing modalities provide reliable burn severity classification, while thermal imaging achieved higher standalone classification performance. The proposed synchronized dataset and controlled experimental framework provide an objective benchmark for future multimodal fusion methods for burn severity assessment.