Refinement of an Artificial Intelligence Algorithm for Enhanced Burn Wound Depth Assessment Using Multispectral Imaging: An Expanded Proof of Concept Study.

Background: With the advent of Convolutional Neural Networks (CNNs), artificial intelligence is now applicable to visual fields. We used multispectral imaging (MSI) sensors capable of detecting wavelengths outside visible spectra to image burn wounds. The output was converted to pixel-level data and analyzed by an array of CNNs to inform development of a Deep Learning (DL) algorithm for burn assessment.

Methods: Three burn centers prospectively grouped consenting subjects into those with wounds likely to heal nonoperatively by 21 days, or those benefiting from surgery. Both groups underwent MSI sensor imaging at enrollment and once daily until discharge/excision. Nonoperative subjects were evaluated at 21 days, while operative subjects underwent biopsies. A "Truthing Panel" of burn experts created a "ground truth" for each wound that was converted to pixel-level data and used to train ten CNNs (eight unique DL algorithms and two ensemble DL algorithms).

Results: 1037 MSI images and 161 biopsies were collected from 100 adult and 24 pediatric subjects. The most effective CNN algorithm exhibited an Area Under the Curve of 0.95 (accuracy= 89.29%, sensitivity= 90.51%, specificity= 87.22%) with the covariate "time-since-injury" found to be significant (p < 0.0001). Accuracy was lowest, 88.5%, at 1 - 2 days after injury and highest, 93.5%, at 3 - 4 days. The CNN's learning curve predicted an accuracy of 94.04% after enrolling 374 subjects in a future training study.

Conclusions: An optimal CNN architecture and the importance of "time-since-injury" as a covariate were identified, informing the design/powering of upcoming algorithm Training and Validation Studies.