Impact of Imaging Protocols on Thermal Detection of Pressure Injuries: Threshold versus Deep Learning Across Skin Tones

This study demonstrates that deep learning models significantly outperform traditional threshold-based approaches in detecting simulated pressure injuries via thermal imaging, achieving robust and equitable accuracy across diverse skin tones and varying imaging protocols by leveraging spatial thermal gradients rather than fixed temperature differentials.

The Gist

The Big Picture: Finding the "Cold Spot" Before the Bruise Shows

Imagine you have a bruise forming under your skin. Before you can see the purple or red mark, the area underneath actually gets cooler because blood flow is slowing down. Doctors have long known this, and they use special thermal cameras (like night-vision goggles for heat) to try and spot these "cold spots" early to prevent serious bedsores (pressure injuries).

However, there's a problem: Skin comes in many colors, and cameras come in many types. The big question this paper asked was: Does the method we use to find these cold spots work equally well for everyone, regardless of their skin tone or the camera we use?

The Two Detectives: The "Rulebook" vs. The "Smart Learner"

The researchers tested two different ways to analyze the thermal images:

1. The Rulebook Detective (Threshold-Based Approach):

   • How it works: This method follows a strict, simple rule. It measures the temperature of the "cold spot" and compares it to a "normal spot" nearby. If the difference is bigger than a specific number (like -1.71°C), it screams "Alert! Bedsores!" If not, it says "All clear."
   • The Analogy: Think of this like a security guard at a club who only lets people in if they are taller than 6 feet. It's a single, rigid rule. It doesn't matter if the person is wearing a hat, standing on a box, or if the lighting is dim; if they aren't 6 feet tall, they don't get in.

2. The Smart Learner (Deep Learning/CNN Models):

   • How it works: Instead of a single number, this method uses Artificial Intelligence (AI) to look at the whole picture. It learns to recognize the shape, the edges, and the pattern of how the heat fades away from the cold spot.
   • The Analogy: Think of this like a seasoned art critic. They don't just measure the height of a painting; they look at the brushstrokes, the lighting, the composition, and the overall vibe. They understand the context of the image, not just one specific measurement.

The Experiment: A Controlled "Cool-Down"

To test these detectives, the researchers didn't wait for real bedsores to form (which takes days). Instead, they created a safe, controlled simulation:

• The Subjects: 35 healthy adults with a wide variety of skin tones (from very light to very dark).
• The Trick: They placed a cold stone cylinder on a specific spot on the participants' lower backs for 5 minutes to simulate the cooling effect of a developing pressure injury.
• The Variables: They took pictures using two different cameras (a high-end professional one and a cheaper, lower-resolution one) under 12 different conditions (different lights, different distances, different body positions).

The Results: Who Won the Race?

1. The Smart Learner (AI) Crushed the Rulebook
The AI models were much more accurate (about 99% accuracy) compared to the Rulebook method (about 95.6% accuracy).

• Why? The Rulebook is too rigid. If the camera is slightly different or the lighting changes, the "magic number" for the temperature difference gets messed up.
• The Camera Problem: When the researchers used the cheaper, lower-resolution camera, the Rulebook detective got confused and made many more mistakes, especially on people with medium-dark skin tones. The Smart Learner, however, stayed calm and accurate on both cameras.

2. The Skin Tone Surprise
The Rulebook method was unfair. It struggled the most with people who had medium-dark skin tones (MST 6) when using the cheaper camera. It also struggled with the darkest skin tones on the expensive camera.

• The AI Advantage: The Smart Learner treated everyone fairly. It performed consistently well across all skin tones, proving that it wasn't biased by how much melanin was in the skin.

3. What Was the AI Actually Looking At?
The researchers used a special tool (Grad-CAM) to see where the AI was "looking" in the images.

• The Discovery: The AI wasn't just looking at the center of the cold spot. It was focusing on the edges or the boundaries where the cold area meets the warm skin.
• The Analogy: Imagine a snowball melting on a warm sidewalk. The Rulebook just checks the temperature of the center of the snowball. The AI looks at the crunchy edge where the snow is turning to water. The AI realized that the shape of the temperature change is what matters, not just the temperature itself.

Why Did the AI Sometimes Fail?

Even the Smart Learner made a few mistakes. The paper found that these mistakes happened when the "cold spot" started to warm up again (rewarming).

• The Confusion: As the cold spot warmed up, the sharp edge between cold and warm started to blur. The AI got distracted by other warm spots on the body (like near the spine) and lost focus on the original cold spot.
• The Lesson: This suggests that the AI is very good at spotting the pattern of cooling, but if the pattern gets too fuzzy or weak, it can get confused.

The Bottom Line

This paper shows that for detecting early signs of pressure injuries using thermal cameras:

• Don't rely on a single temperature number. It's too fragile and depends too much on the camera and the person's skin.
• Use AI that looks at the whole picture. By understanding the shape and gradients of the heat, AI can be fair and accurate for people of all skin tones, even when using different cameras or taking photos in less-than-perfect conditions.

The study concludes that while we need more testing on real patients, the "Smart Learner" approach is a much more promising and equitable tool for the future of healthcare than the old "Rulebook" method.

Technical Summary

Technical Summary: Impact of Imaging Protocols on Thermal Detection of Pressure Injuries

Problem Statement
Pressure injuries represent a significant healthcare challenge, causing approximately 60,000 deaths annually in the United States and costing over $26.8 billion. Early detection is critical but difficult; visual inspection is unreliable, particularly for patients with darker skin tones where non-blanchable erythema is hard to identify. While thermal imaging offers a non-contact method to detect early temperature changes associated with tissue ischemia, its clinical utility is hindered by a lack of understanding regarding how imaging protocol variations (lighting, distance, positioning, camera type) and diverse patient skin tones affect detection accuracy. Furthermore, current clinical approaches often rely on fixed temperature thresholds, which may not generalize across different hardware or physiological populations.

Methodology
This study utilized a controlled simulation to evaluate two classification approaches for detecting thermal cooling, a proxy for early pressure injury temperature changes.

• Data Collection: The researchers collected 1,680 thermal images from 35 healthy adults with diverse skin tones, stratified by the Monk Skin Tone (MST) scale into three groups: MST 1-5 (n=5), MST 6 (n=7), and MST 7-10 (n=23).
• Experimental Protocol: A controlled cooling protocol was induced by placing a cooled stone cylinder on the posterior superior iliac spine (PSIS) for five minutes. Images were captured using two distinct cameras: a high-resolution FLIR E8XT (320×240 thermal resolution) and a consumer-grade FLIR ONE Pro (160×120 thermal resolution).
• Imaging Variations: A factorial design was employed with 12 imaging configurations varying in lighting (ambient vs. ring light), distance (35 cm vs. 50 cm), and posture (forward, stacked, and backward knee placement).
• Classification Approaches:
  1. Threshold-Based: A fixed temperature differential threshold of -1.71°C (derived via ROC analysis) was applied to classify images as "cooling" or "baseline" based on the temperature difference between the cooled region and a control region.
  2. Deep Learning (CNN): Three transfer learning models (MobileNetV2, InceptionNetV3, ResNet50) were fine-tuned to classify images. The models utilized per-image normalization, data augmentation (flips, rotation), and a stratified participant-level 5-fold cross-validation framework to prevent data leakage.
• Explainability: Gradient-weighted Class Activation Mapping (Grad-CAM) was used to visualize the spatial features influencing CNN predictions.

Key Results

• Overall Performance: Deep learning models significantly outperformed the threshold-based approach. CNNs achieved overall accuracies ranging from 98.6% to 99.6%, whereas the threshold-based method achieved 95.6%.
• Robustness Across Protocols: The threshold-based approach exhibited protocol-dependent misclassification patterns, with error rates varying from 1.9% to 13.0% depending on lighting and distance. In contrast, CNNs maintained consistent performance across all 12 imaging protocols, with four protocols achieving zero misclassifications.
• Hardware Dependency: The threshold-based method showed significant hardware dependency. On the low-resolution FLIR ONE Pro, its error rate jumped to 5.2% (compared to 1.2% on the high-resolution E8XT), with specific vulnerability in the intermediate skin tone group (MST 6), which accounted for 22 of 44 total misclassifications on this device. CNNs demonstrated superior robustness to camera type, maintaining low error rates (0.8% on FLIR ONE Pro) across all skin tones.
• Skin Tone Equity: While the threshold-based approach showed varying performance across skin tone groups (lowest accuracy in MST 6), CNNs maintained high accuracy (97.5%–99.7%) across all groups.
• Error Analysis: Grad-CAM visualizations revealed that CNNs focused on thermal gradients at the boundaries of the cooling region. Misclassifications occurred primarily when thermal gradients diminished over time due to rewarming, causing the model's attention to shift toward competing warm artifacts (e.g., paraspinal warming) rather than the cooling region itself. This failure mode was consistent across skin tones, suggesting the errors were driven by temporal thermal dynamics rather than skin tone physiology.

Significance and Claims
The paper claims that deep learning-based approaches offer a more equitable and robust solution for thermal imaging-based pressure injury detection compared to fixed threshold methods. Key contributions include:

1. Superior Generalization: CNNs preserve spatial temperature information and thermal gradients, allowing them to generalize across varying imaging conditions and camera types where fixed thresholds fail.
2. Equity Across Skin Tones: Unlike threshold-based methods, which showed camera-dependent vulnerability patterns across different skin tones, deep learning models maintained consistent performance, suggesting they are less susceptible to the physiological and measurement variations that affect single-value thresholds.
3. Clinical Implications: The findings suggest that clinical implementation of thermal imaging does not require rigid standardization of lighting, distance, or positioning, as CNNs can tolerate natural variations. Furthermore, the study argues that datasets for training such systems should be labeled based on clinical ground truth (confirmed injury status) rather than post-hoc temperature thresholds, which may not reflect the complex spatial dynamics of tissue injury.

The authors conclude that while CNNs show promise for reliable detection across diverse populations, future validation in clinical populations with actual pressure injuries and larger, balanced cohorts is necessary to confirm these technical advantages translate to improved early detection in real-world settings.