Speaker
Description
Active infrared thermography (IRT) offers a non-invasive route to detect and characterise
cutaneous lesions by measuring transient surface temperature responses following controlled
cooling. This thesis presents an end-to-end investigation combining physics-based
simulation, thermogram creation, algorithmic feature extraction, machine learning and
prototype instrumentation to assess the feasibility of lesion localisation and parameter
estimation (diameter, depth, shape) from reheating sequences.
A five-layer parametric finite-element skin model based on Pennes’ bioheat equation was
developed to produce large, labelled synthetic datasets. Data pipelines converted the
generated 3D thermal data into clinically comparable 2D thermograms. Three network
architectures were evaluated — a CNN regressor (diameter/depth), a U-Net segmentation
network (localisation/area) and a 3D-CNN for spatiotemporal reheating analysis.
Experimental validation used skin-mimicking phantoms to characterise the cooling
method and the HypIRskin prototype; a multi-camera imaging platform was developed
as a scalable pathway for clinical acquisition.
Under ideal, noise-free simulation conditions deterministic methods recovered lesion
metrics with sub-millimetric error (depth ≤0.05mm; diameter/shape ≤0.1 mm). Machine learning
models trained on synthetic data achieved practical accuracies (regression: diameter
≈0.2 mm, depth ≈0.12 mm; segmentation reliable for lesions ≥1 mm). The
HypIRskin prototype produced repeatable reheating curves within an operational timeframe
(≈3 minutes per lesion). Limitations include reliance on synthetic data, limited
clinical samples and sensitivity to cooling protocol, calibration and registration.
These results indicate that active IRT preserves measurable surface signatures that reflect
underlying lesion geometry, and that physics-based simulation can reliably generate
labeled training data when clinical examples are limited. Successful translation will
require a large, carefully annotated clinical dataset to enable robust validation.
