AI-Aided Spectroscopy Assesses Burn Severity to Improve Recoveries

Researchers at Stony Brook University developed a neural network model that uses terahertz time-domain spectroscopy (THz-TDS) to evaluate burns and predict healing outcomes. The latest work builds on a previous advancement in which the researchers developed a portable, hand-held imaging device for fast THz-TDS imaging of burn injuries, which aims to make the method practical for clinical use.

Physical changes caused by a burn will produce alterations in the THz reflectivity. For this reason, THz-TDS, which uses short pulses of THz radiation to probe a sample, shows potential as a tool for accurately estimating burn severity.

The initial assessment of the depth of a burn injury forms the basis for the course of treatment. However, severe skin burns are highly dynamic and are therefore hard to assess and predict. According to professor M. Hassan Arbab, current methods of burn depth evaluation, which rely on visual and tactile examination, have been shown to be unreliable. For these methods, Arbab said, accuracy rates range from 60% to 75%.

Researchers developed a new, physics-based neural network model that uses terahertz time-domain spectroscopy (THz-TDS) data for noninvasive burn assessment. They combined the new approach with the hand-held imaging device shown, which was developed for fast THz-TDS imaging of burn injuries. Courtesy of Stony Brook University.

In their earlier research, team members used numerical methods to extract features from the THz-TDS images and machine learning techniques to estimate the severity grade of in vivo burn injuries using measurements obtained from the PHASR scanner — the portable device developed by the team. However, this approach did not consider the physical dynamics and macroscopic changes of the dielectric permittivity of burned skin tissue.

Dielectric permittivity describes how a material responds to an electric field; the way burned skin tissue responds to an electrical field could indicate changes in the burn’s severity. The physical dynamics and macroscopic changes of the dielectric functions of burned tissue can alter the tissue’s THz reflectivity and are thought to be responsible for the contrast in THz-TDS measurements of burns with different depths of dielectric permittivity.

The researchers turned to the double Debye theory, which has previously been used to explain the interaction of THz radiation with biological tissue. The researchers used the theory to model the permittivity of burned tissue. They developed a methodology for calculating the physical THz spectroscopic features of in vivo burn injuries using double Debye parameters.

“We developed a neural network model utilizing the five parameters obtained from fitting the double Debye model to the dielectric permittivity of burn injuries,” Arbab said.

The researchers demonstrated that the five parameters of the double Debye model could form an artificial neural network classification algorithm capable of automatic diagnosis of the severity of burn injuries, and was also capable of predicting the re-epithelialization status of the wound 28 days post-burn.

A THz spectroscopic image of a burn obtained with the PHASR scanner by sweeping over a frequency range between 0.2 and 0.9 THz. Courtesy of Stony Brook University.

Use of THz-TDS for diagnosing the severity of burn injuries was initially limited to point-spectroscopy measurements, which did not account for the heterogeneity and the spatial variations in burns. Further, though various other technologies have been developed to improve burn assessment, they too have failed to achieve wide adoption due to other drawbacks, such as as long acquisition times, high costs, and limited penetration depth and field of view. Previous THz spectroscopy setups, though technologically promising, were bulky and expensive, making them unsuitable for clinical use.

The researchers developed the portable hand-held spectral reflection (PHASR) scanner, which enables fast hyperspectral imaging of in vivo burn injuries using THz-TDS, to address this issue.

“This hand-held device uses a dual-fiber femtosecond laser with a center wavelength of 1560 nm and terahertz photoconductive antennas in a telecentric imaging configuration for the rapid imaging of a 37- × 27-mm² field of view in just a few seconds,” Arbab said.

The researchers also implemented numerical feature extraction and machine learning techniques to automatically estimate the severity grade of in vivo burn injuries using the PHASR scanner measurements. They obtained spectroscopic images of in vivo porcine skin burns and measured the permittivity of the burns using the PHASR scanner. After determining the Debye parameters, the researchers used this data to create a neural network model based on labeled biopsies. The Debye parameters provided a feature extraction strategy based on the dielectric function of the tissue, instead of using purely numerical features.

According to Arbab, the physics-based approach helps to ultimately reduce the dimensionality of THz data for training the AI models. This, he said, improves the efficiency of the machine learning algorithms.

The Debye parameters when combined with the neural network model were able to predict the outcome of the wound healing process with an accuracy rate of 93% and the severity group of the burns with an average accuracy rate of 84.5%.

Because the new burn assessment technique reduces the number of input variables of the neural network models to only five Debye parameters, it lowers the amount of training data necessary by at least two orders of magnitude.

The researchers said that clinical testing of both the technique and the hand-held imaging device are needed before the technique can be integrated into the existing workflow of clinical burn assessment. However, the reported accuracy rates promise a robust strategy for diagnosis of burn injuries.

The research was published in Biomedical Optics Express (www.doi.org/10.1364/BOE.479567).