# Research on a Burn Severity Detection Method Based on Hyperspectral Imaging

## TL;DR

This paper introduces a new method using hyperspectral imaging and a neural network to accurately detect the severity of burn wounds.

## Contribution

The novel MBNet model improves burn severity detection by capturing spectral feature dependencies with a bidirectional scanning strategy.

## Key findings

- MBNet outperforms seven machine learning algorithms in burn severity classification accuracy.
- HSI combined with MBNet effectively captures subtle spectral differences in burn-affected skin.
- The proposed method offers a non-invasive and reliable alternative to traditional burn diagnosis.

## Abstract

The accurate detection of burn wounds is a key research direction in the field of burn medicine, as diagnostic results directly influence the risk of wound infection and the formation of hypertrophic scars. Currently, burn diagnosis is primarily dependent on the clinical judgment of physicians, but its accuracy is typically only between 65% and 70%. Therefore, a non-invasive, efficient method for burn severity assessment is urgently needed. Hyperspectral imaging (HSI), as a non-invasive and contactless spectral detection technique, has been shown to precisely monitor structural changes in burn-affected skin tissue and holds significant potential for burn depth diagnosis. However, research on the application of burn severity detection remains relatively limited, which restricts its widespread use in clinical settings. A burn severity detection classification network (MBNet) based on the Mamba model is proposed in this paper. Through a bidirectional scanning strategy, MBNet effectively captures the long-term dependencies of spectral features, accurately establishes the relationships between bands, and efficiently distinguishes subtle spectral differences under different burn conditions. MBNet provides a reliable and efficient method for clinical burn severity assessment. A comparison of MBNet with seven typical machine learning algorithms on a custom dataset demonstrates that MBNet significantly outperforms these methods in terms of accuracy.