Miriam Asare-Baiden,Kathleen Jordan,Andrew Chung,Sharon Eve Sonenblum,Joyce C. Ho
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.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: whether thermal imaging technology can be a feasible method for detecting pressure injury (PI) in patients with dark skin. Specifically, the paper explores whether thermal imaging technology can effectively detect the temperature changes of patients with dark skin under different lighting conditions, camera distances, patient postures, and camera resolutions, so as to predict early - stage pressure injuries.
### Background and Motivation
Pressure injury (PI) refers to "local damage to the skin and subcutaneous soft tissue, usually occurring at bony prominences or related to medical devices" (European Pressure Ulcer Advisory Panel et al., 2019). PI is a major medical problem, and at least 10% of hospitalized adults will develop PI (Li et al., 2020). In addition, PI is associated with a decline in quality of life (Khor et al., 2014), an increase in mortality (Bergquist - Beringer et al., 2013), a prolongation of hospital stay, and an increased likelihood of requiring institutional care after discharge.
Currently, methods for assessing PI risk include using the Braden scale (Bergstrom et al., 1987; Borghardt et al., 2015) and visual inspection. Visual inspection includes evaluating skin temperature changes, changes in tissue consistency, and the presence of tenderness. However, for patients with dark skin, visual inspection is particularly difficult, resulting in ethnic differences in PI outcomes (Black et al., 2023). Given that PI may lead to partial or complete capillary occlusion, thermal imaging technology has been proposed as an alternative technique for PI detection (Baron et al., 2023).
### Limitations of Existing Research
Although existing research shows that deep - learning models such as convolutional neural networks (CNN) exhibit high predictive performance in detecting PI (Wang et al., 2021; Pandey et al., 2022; Fergus et al., 2023), these studies have the following three main limitations:
1. **Insufficient verification for different skin colors**: There are few existing studies on the application of thermal imaging on different skin colors (Aloweni et al., 2019; Sprigle et al., 2003). Infrared thermometers underestimate fever in African - Americans (Bhavani et al., 2022), but research on the sensitivity of thermal imaging to melanin is still insufficient in PI detection.
2. **Lack of comparison between thermal imaging and optical imaging**: No study has compared the effectiveness of thermal imaging and optical imaging in PI detection.
3. **Strict data collection protocols**: The data collection protocols in existing research are very strict, such as fixed lighting, camera distance, and angle, which are impractical in the actual clinical environment.
### Research Objectives
This study aims to address the above limitations in the following ways:
- **Explore the feasibility of thermal imaging in patients with dark skin**: Detect temperature changes under different patient postures, lighting conditions, camera distances, and camera resolutions.
- **Collect a diverse data set**: Collect optical and thermal imaging data from 35 healthy adults (mainly with dark skin), and induce temperature changes through cooling and cupping protocols.
- **Evaluate model performance**: Use a convolutional neural network (CNN) model to evaluate its performance in detecting temperature changes.
### Methods
#### Data Collection
- **Participants**: 35 healthy adults, of which 30 have relatively dark skin (Monk Skin Tone Scale level 6 and above), and 5 have relatively light skin (Monk Skin Tone Scale level 5 and below).
- **Imaging Devices**: Use two thermal imaging cameras, FLIR E8 - XT (320x240 resolution) and FLIR Pro One (160x120 resolution).
- **Imaging Protocols**: Include two lighting settings (ambient light and ring light), two distances (35 cm and 50 cm), and three postures (front knee, crossed knee, back knee).
#### Model and Task
- **CNN Model**: Use MobileNetV2, which is composed of...