📄 radiology and imaging

Hyperspectral imaging in the emergency department to characterize lower leg edema

This study demonstrates that visible/near-infrared hyperspectral imaging, particularly when analyzed using full-spectrum machine learning algorithms, can accurately and rapidly classify cellulitis, non-cellulitis edema, and healthy lower-leg tissue in emergency department patients across diverse skin pigmentation types.

Original authors: Kasi, K., May, C., Gallegos, H., Kobayashi, L., Conway, B. R., Pare, J. R.

Published 2026-09-16

📖 4 min read☕ Coffee break read

CC0 1.0

Original authors: Kasi, K., May, C., Gallegos, H., Kobayashi, L., Conway, B. R., Pare, J. R.

Original paper dedicated to the public domain under CC0 1.0 (https://creativecommons.org/publicdomain/zero/1.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

When a doctor looks at a swollen leg in an emergency room, they are often trying to solve a difficult puzzle. The swelling could be a harmless buildup of fluid, or it could be a serious bacterial infection called cellulitis that requires immediate antibiotics. Sometimes the signs are so similar that even experienced clinicians struggle to tell them apart, leading to delays in treatment or unnecessary medication. To help solve this, researchers are exploring a way to see inside the skin without cutting or probing. This approach uses a special kind of camera that captures light far beyond what the human eye can see. While our eyes see only red, green, and blue, these cameras record hundreds of narrow bands of light, stretching from the visible colors into the invisible near-infrared range. When light hits the skin, it bounces back carrying a unique signature shaped by the blood, water, and other chemicals living just beneath the surface. By analyzing these light signatures, scientists hope to find a reliable, instant way to distinguish between different types of leg swelling.

A team of researchers recently put this idea to the test in a busy emergency department. They set up a hyperspectral camera to photograph the lower legs of 83 patients who had come in with swelling. The goal was to see if the light patterns could accurately sort the patients into three groups: those with a healthy leg, those with non-infectious fluid buildup, and those with the bacterial infection. The team compared two ways of reading the data. The first method relied on standard, simplified measurements that doctors already use, such as estimates of oxygen levels and blood volume. The second method looked at the entire, complex spectrum of light, letting a computer find patterns across all the hundreds of colors at once. They tested these methods on patients with different skin tones to ensure the technology worked fairly for everyone.

The results showed that the full-spectrum approach was a powerful tool, but its strength depended on what the doctors were trying to diagnose. When the goal was to spot the bacterial infection, both the standard measurements and the full-spectrum analysis worked very well, correctly identifying the disease in most cases. This suggests that the infection creates such strong changes in blood flow and oxygen that even simple tools can find it. However, the story changed when the researchers tried to distinguish between healthy legs and those with non-infectious swelling. In this case, the standard measurements failed, performing no better than random guessing. The full-spectrum analysis, by contrast, succeeded, correctly identifying the fluid buildup in a significant majority of cases. This finding indicates that non-infectious swelling leaves a subtle, complex fingerprint in the light that simple measurements miss, but that a detailed look at the entire light spectrum can reveal.

The study also addressed a critical concern about fairness in medical technology. Because darker skin absorbs more light, there is a fear that new imaging tools might work poorly for people with higher levels of melanin. The researchers carefully checked whether the accuracy of their system changed based on the patients' skin tone. They found no evidence that the system performed differently for different skin types within the range of people they studied. The technology appeared to work equally well for all patients, suggesting that the light-based method could be a reliable tool for everyone, regardless of pigmentation.

While the results are promising, the researchers are careful to note that this is a step forward, not a final solution. The study was conducted in a real-world emergency setting with all the usual challenges of lighting and patient movement, which makes the success of the technology particularly encouraging. The team found that looking at the whole picture of light data provided more useful information than breaking it down into a few simple numbers, especially for the harder-to-diagnose conditions. They also discovered that summarizing the data for the whole leg was more stable than looking at individual tiny spots of skin, likely because it smoothed out the natural variations found in any single patch of tissue. These findings suggest that hyperspectral imaging could eventually become a standard, non-invasive tool in emergency rooms, helping doctors make faster and more accurate decisions for patients with swollen legs.

Technical Summary: Hyperspectral Imaging in the Emergency Department to Characterize Lower Leg Edema

Problem Statement
Differentiating cellulitis (a bacterial skin infection) from non-infectious lower-leg edema in the emergency department (ED) is a significant clinical challenge. These conditions often present with similar visual signs but require distinct management strategies; misdiagnosis occurs in approximately 30% of cases, leading to either delayed treatment or unnecessary antibiotic use. While Hyperspectral Imaging (HSI) offers a non-invasive method to quantify tissue chromophores (such as hemoglobin, water, and melanin) beyond standard RGB imaging, its reliability in the variable, uncontrolled conditions of an ED remains uncertain. Furthermore, it is unclear whether reducing high-dimensional spectral data to standard physiological indices (e.g., tissue oxygenation, hemoglobin concentration) sacrifices diagnostic information, particularly regarding skin pigmentation, which can obscure erythema and bias model performance.

Methodology
The study enrolled 83 patients from a Level 1 trauma center ED, comprising patients with cellulitis ( $n=18$ ), non-cellulitis edema ( $n=43$ ), and healthy controls ( $n=42$ ). Data was acquired using a Surface Optics Corporation SOC-710-VP hyperspectral camera capturing 128 spectral bands (377–1043 nm) under tungsten illumination. The acquisition protocol included three camera configurations (left leg, right leg, and both legs) to accommodate various patient positions.

Preprocessing: Raw data was calibrated against a Spectralon reference and a dark frame. Analyses were restricted to the 400–1000 nm range. To mitigate variations in geometry, curvature, and ambient light, Standard Normal Variate (SNV) normalization was applied to each spectrum.
Data Labeling: Labels were assigned at the leg level based on physician diagnosis. Pixels from affected legs were labeled as cellulitis or non-cellulitis edema; unaffected contralateral legs were labeled as healthy.
Classification Approaches: The study compared two analytical strategies:
1. Standard Indices: Four spectral indices were calculated: Tissue Oxygenation ( $StO_2$ ), Total Hemoglobin Index (THI), Tissue Water Index (TWI), and Near-Infrared Perfusion (NIR). These were fed into logistic regression models.
2. Whole-Spectrum Analysis: Full spectral data (400–1000 nm) was analyzed using Partial Least Squares Discriminant Analysis (PLS-DA).
Validation: Models were evaluated using Leave-One-Group-Out (LOGO) cross-validation, where all data from a single patient was held out for testing to prevent data leakage. Performance was assessed at both the patient level (using median spectra per patient) and the pixel level.
Sensitivity Analysis: The impact of Fitzpatrick skin type (Types 1–5 represented) on classification accuracy was tested using logistic regression to determine if pigmentation drove model performance.

Key Results

Cellulitis vs. Healthy: Both whole-spectrum models and standard indices performed robustly, achieving high Area Under the Curve (AUC) values (Patient-level: 0.93 for whole-spectrum vs. 0.96 for indices). This suggests that standard indices capturing oxygenation and perfusion are sufficient for distinguishing cellulitis from healthy tissue.
Non-Cellulitis Edema vs. Cellulitis: Both approaches performed similarly well (Patient-level AUC: 0.92 for whole-spectrum vs. 0.89 for indices).
Non-Cellulitis Edema vs. Healthy: This comparison revealed the most significant divergence. Standard indices performed poorly (Patient-level AUC: 0.41), whereas whole-spectrum analysis substantially improved discrimination (Patient-level AUC: 0.72; Pixel-level AUC: 0.78 vs. 0.50 for indices). This indicates that distinguishing non-infectious edema from healthy tissue relies on spectral information distributed across the full spectrum rather than predefined physiological indices.
Skin Pigmentation: Sensitivity analysis found no significant association between Fitzpatrick skin type and classification accuracy for any model. The models did not appear to be driven primarily by pigmentation differences within the tested range (Types 1–5).
Resolution: Patient-level median models generally provided more stable performance than pixel-level models, likely due to the reduction of local spatial variability and noise.

Significance and Claims
The paper claims that visible/near-infrared HSI acquired in a real-world emergency department setting can effectively distinguish between cellulitis, non-cellulitis edema, and healthy tissue. The primary contribution is the demonstration that while standard spectral indices are adequate for detecting the inflammatory signals of cellulitis, they fail to capture the subtle optical changes associated with non-infectious edema. The study argues that retaining full-spectrum information is necessary to recover diagnostic data lost through spectral reduction, particularly for differentiating edema from healthy tissue.

Furthermore, the authors claim that their SNV-normalized approach yields models that are robust to skin pigmentation within the studied cohort, suggesting potential for equitable diagnostic performance across different skin types. The paper concludes that HSI is a viable point-of-care tool and that full-spectrum analysis serves as a valuable complement to standard clinical indices, warranting further evaluation in larger and more diverse populations.

More like this