Abstract
The severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) pandemic rapidly spread across the world, creating a need for real-time screening of individuals to prevent further transmission. We aimed to develop a mobile thermographic imaging platform, intended to scan an entrance or hallway to identify people with abnormal facial temperatures and/or vital signs. Adult patients who were actively symptomatic with SARS-CoV-2 infection were included along with people with other febrile illnesses and healthy individuals. We applied an attention-based model to conduct a thermal image-based diagnosis. While the study faced challenges with model convergence, it provides critical insights into the limitations of thermal imaging for clinical screening.
Introduction
During the height of the SARS-CoV-2 pandemic, non-invasive, high-throughput screening methods for public spaces like airports and hospitals were urgently needed. Traditional manual temperature checks are slow and prone to error. This research, published in Academic Dermatology, investigates whether a mobile thermographic imaging platform could identify "thermal signatures"—including abnormal facial temperatures, respiratory rates, and heart rates—unique to COVID-19. The goal was to move beyond simple threshold-based checks toward a more holistic, automated diagnostic tool using computer vision.
Methods
The study employed a multi-disciplinary approach combining clinical data collection with advanced computer vision.
Data Acquisition and Participants
Clinical investigators used hand-held thermal cameras to record 2–5 minutes of video of 76 participants at distances of 5 and 10 feet. The cohort included 35 SARS-CoV-2 positive patients, 21 with other infectious diseases (e.g., pneumonia), and 20 asymptomatic controls. A "black body" reference heat source set at 37°C was used for calibration.
AI Architecture and Preprocessing
The team utilized a Time-Space Transformer (TimeSformer). Unlike standard Convolutional Neural Networks (CNNs), Transformers use attention mechanisms to capture long-range dependencies across both space (within a frame) and time (across the video sequence). From each video, 15 frames were sampled at equal intervals and resized to 224x224 pixels for model input.
Results
The primary quantitative result was the failure of the model to converge. The loss function did not reach a stable solution, indicating the neural network could not reliably distinguish COVID-positive patients from the control group.
Statistical Baseline
Statistical analysis revealed significant overlap in vital signs between groups:
- Temperature: The mean temperature for the COVID group was 98.0°F vs. 97.9°F for controls (p=0.82).
- Respiratory Rate: This metric approached significance (17.17 breaths/min for COVID vs. 15.9 for controls, p=0.07), suggesting it may be a more viable marker than temperature.
- Other Vitals: Heart rates and oxygen saturation levels showed no statistically significant differences.
Discussion: Lessons Learned
The study identified several critical "noise" factors that hindered the AI's success:
- Anti-Pyretic Confound: Many symptomatic patients were taking fever-reducing medications, effectively masking the thermal signal.
- Hardware Instability: Hand-held cameras introduced "camera shake," creating spatial noise that interfered with pixel-level vital sign calculations.
- Internal Processing: Thermal cameras often use "black box" auto-adjustment algorithms that can distort raw data needed for precise training.
Conclusion
This study provides a valuable baseline for future infrared-based diagnostic research. Future work should focus on anomaly detection using autoencoders trained on healthy populations and utilize stabilized, tripod-mounted hardware to ensure data purity.
Summary updated