Is thermography a viable solution for detecting pressure injuries in dark skin patients?

Miriam Asare-Baiden, Kathleen Jordan, Andrew Chung, Sharon Eve Sonenblum, Joyce C. Ho

Abstract

Visual inspection for pressure injuries (PIs) is notoriously unreliable for patients with dark skin, leading to significant healthcare disparities. This paper demonstrates that thermography (thermal imaging) combined with deep learning offers a robust, skin-tone-agnostic solution that outperforms traditional optical imaging. The study shows that thermal models remain accurate regardless of lighting, camera distance, or patient posture, providing a potential path toward more equitable clinical diagnostics.

Introduction

Pressure injuries, commonly known as bedsores, represent localized damage to the skin and underlying tissue. Early detection is critical; identifying "erythema" (the redness associated with inflammation) allows clinicians to intervene before the skin breaks. However, visual inspection—the current standard of care—fails significantly for patients with high melanin levels. Because redness is harder to discern on darker skin, these patients are often diagnosed later, leading to higher mortality rates and longer hospital stays.

While Convolutional Neural Networks (CNNs) have been proposed for PI detection, most models are trained on datasets lacking skin tone diversity and rely on "perfect" lab conditions. This study investigates whether thermal imaging can bridge this gap and remain "rugged" enough for real-world clinical use.

Methods and Materials

To evaluate the efficacy of thermal imaging, researchers simulated the physiological markers of PIs in a controlled study of 35 healthy adults, predominantly with dark skin tones. Two protocols were used to induce detectable skin changes: cooling (applying a cooled stone cylinder) and cupping (using suction to induce erythema/inflammation).

Images were captured using two different thermal cameras (FLIR E8-XT and FLIR Pro One) across 12 protocols that varied environmental factors, including lighting (ambient vs. ring light), distance (35cm vs. 50cm), and patient posture. For classification, the researchers employed MobileNetV2, a lightweight CNN architecture. The model was trained on three data types: standard optical images, grayscale thermal images, and colorized thermal images, with the goal of binary classification between affected spots and control skin.

The Dataset

The study introduced a new, diverse dataset specifically focused on darker skin tones:

  • Participants: 35 individuals categorized by the Eumelanin Human Skin Colour Scale (Intermediate Low to Intermediate High).
  • Cooling Dataset: 1,680 images.
  • Erythema Dataset: 314 images captured within 7 minutes of cupping.

Results and Discussion

Comparison of optical and thermal imaging across various dark skin tones (Intermediate Low to Intermediate High).

Thermal Imaging as a "Skin-Tone Blind" Solution

A primary finding of the study is that thermal imaging is largely indifferent to melanin levels. While melanin absorbs infrared radiation, it does not mask the thermal signals necessary for PI detection. Thermal models achieved near-perfect performance across all skin tones, whereas optical models struggled as melanin levels increased. In the cooling task, thermal imaging (both B&W and Color) achieved a perfect 1.0 AUC and F1-score, while optical imaging lagged at 0.818 AUC.

Robustness to Clinical Chaos

Thermal imaging demonstrated significant resilience to environmental noise. Optical models were highly sensitive to specific combinations of lighting and pose—for instance, being easily tripped up by shadows from a ring light. In contrast, the thermal signal remained clear regardless of these variations. Interestingly, for optical images, a greater distance (50cm) actually improved performance by reducing the impact of harsh, localized shadows present at 35cm.

Detecting Erythema

Detecting erythema (inflammation) proved more challenging than detecting cooling. Erythema fades quickly and produces more subtle temperature changes, making it a more difficult task for the AI. Nevertheless, thermal imaging remained superior, with the color thermal model reaching an AUC of 0.935 and F1 of 0.918, compared to the optical model's AUC of 0.909.

Limitations and Future Work

While promising, this preliminary work has several limitations:

  • Simulated Markers: The study used cooling and cupping to mimic PIs; real injuries involve complex tissue degradation that may present differently.
  • Sample Size and Anatomy: The cohort was limited to 35 participants and focused solely on the lower back. Future validation is required for other high-risk areas like heels or elbows.
  • Ground Truth Correlation: The ground truth for erythema was determined by a colorimeter, but temperature changes and visual redness do not always correlate perfectly, necessitating further research into defining "positive" cases.

Conclusion

Thermography combined with deep learning represents a viable, equitable solution for pressure injury detection in diverse patient populations. By removing the reliance on visual redness, this technology can mitigate the diagnostic delays currently experienced by patients with dark skin tones.

Summary updated

Abstract

Pressure injury (PI) detection is challenging, especially in dark skin tones, due to the unreliability of visual inspection. Thermography has been suggested as a viable alternative as temperature differences in the skin can indicate impending tissue damage. Although deep learning models have demonstrated considerable promise toward reliably detecting PI, the existing work fails to evaluate the performance on darker skin tones and varying data collection protocols. In this paper, we introduce a new thermal and optical imaging dataset of 35 participants focused on darker skin tones where temperature differences are induced through cooling and cupping protocols. We vary the image collection process to include different cameras, lighting, patient pose, and camera distance. We compare the performance of a small convolutional neural network (CNN) trained on either the thermal or the optical images on all skin tones. Our preliminary results suggest that thermography-based CNN is robust to data collection protocols for all skin tones.
BibTeX
@misc{asarebaiden2024is,
  author = {Miriam Asare-Baiden and Kathleen Jordan and Andrew Chung and Sharon Eve Sonenblum and Joyce C. Ho},
  title = {{Is thermography a viable solution for detecting pressure injuries in dark skin patients?}},
  year = {2024},
  eprint = {2411.10627},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2411.10627},
}
Figure