Refinement of an Artificial Intelligence Algorithm for Enhanced Burn Wound Depth Assessment Using Multispectral...

Area of Science:

• Medical imaging
• Artificial intelligence in medicine
• Burn wound assessment

Background:

• Convolutional Neural Networks (CNNs) enable AI in visual fields.
• Multispectral Imaging (MSI) sensors capture wavelengths beyond visible spectra.
• Developing a deep learning (DL) algorithm for burn assessment using MSI data.

Purpose of the Study:

• To develop and validate a deep learning algorithm for burn wound assessment using multispectral imaging.
• To evaluate the accuracy and effectiveness of CNNs in classifying burn wounds for operative or nonoperative healing.
• To identify key covariates, such as time-since-injury, influencing burn assessment outcomes.

Main Methods:

• Prospective study involving 100 adult and 24 pediatric subjects across three burn centers.
• Multispectral imaging (MSI) of burn wounds, with data converted to pixel-level for analysis.
• Training ten CNNs (8 unique DL algorithms and 2 ensemble DL algorithms) using expert-defined ground truth data.

Main Results:

• The most effective CNN achieved an area under the curve of 0.95 (89.29% accuracy, 90.51% sensitivity, 87.22% specificity).
• Time-since-injury was a significant covariate (P < .0001), with accuracy highest at 3-4 days post-injury (93.5%).
• The CNN learning curve predicted 94.04% accuracy with future subject enrollment.

Conclusions:

• An optimal CNN architecture was identified for burn wound assessment.
• "Time-since-injury" is a crucial covariate for improving DL algorithm performance.
• Findings inform the design and power calculations for future algorithm training and validation studies.