Graph-Tabular Latent Fusion for Non-Contact Body Temperature Prediction from Thermal Facial Landmarks.

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Non-contact body temperature prediction using facial thermography faces challenges from pose variations, occlusions, and missing data.
• Existing methods struggle with inter-subject variability and incomplete thermal facial data.

Purpose of the Study:

• To develop a robust framework for accurate body temperature prediction from facial thermal landmarks.
• To address challenges of pose, occlusion, and missing data in thermal imaging for temperature estimation.

Main Methods:

• A graph-tabular latent-representation fusion framework was proposed.
• A Pearson correlation coefficient (PCC)-guided graph modeled thermal dependencies between facial landmarks.
• A tabular representation preserved global temperature patterns, fused with the graph branch at the latent level.

Main Results:

• The fused framework outperformed stand-alone graph and tabular methods, especially under missing not at random (MNAR) conditions.
• The best configuration reduced prediction error by approximately 6% on the TFD68 dataset.
• Statistical testing confirmed significant improvements in most comparisons against baselines under MNAR imputation.

Conclusions:

• The proposed graph-tabular fusion framework enhances the robustness of body temperature prediction from facial thermography.
• Fusion offers improved accuracy at a manageable increase in inference time, providing a flexible approach for thermal landmark-based applications.