A Comparative Study of Deep Learning and Boundary-Based Registration for Thermal Imaging of Human Forearm

Thermal imaging holds potential in medical diagnostics, but long-duration observational experiments, such as brachial artery occlusion, suffer from subjects’ motion, which complicates reliable temperature oscillation analysis. This work investigates the suitability of Deep Learning (DL) models for low-contrast, feature-sparse thermal registration. We examined a global–local model trained under intensity-driven unsupervised learning and a synthetic rigid supervised approach. Both models were evaluated against a rigid-body registration baseline. A comparative study was conducted to analyze the performance differences. In this specific scenario, the traditional rigid body registration achieved more reliable alignment (a Dice similarity coefficient (DSC) of 0.9760) than the DL-based approaches (DSCs of 0.9752 and 0.9746) over 2399 test images.