Influence of thermal contrast and limitations of a deep-learning based estimation of early-stage tumour parameters in different breast shapes using simulated passive and dynamic thermography

Abstract

To enhance diagnostic sensitivity compared to passive thermography, thermal stress can be applied to the breast surface with the temperatures being measured in the thermal recovery phase, a process called dynamic thermography. This study aims to evaluate the limitations of both passive and dynamic thermography in estimating early-stage tumour parameters across different breast shapes and how to improve the results. Three breast models with thermoregulation were solved numerically using COMSOL Multiphysics®. A neural network developed in PyTorch was used to estimate breast tumour location and size. The estimates obtained using each approach were compared, and the effects of thermal contrast, noise, and tumour depth range were analysed. Dynamic thermography provided the most accurate estimates compared to passive thermography, with mean error reductions that reached up to 33.25%. Additionally, the number of estimates with errors higher than 10% was up to 48.42% lower. Tumour radius showed the lowest noise threshold, providing the highest estimations errors. Adding deeper tumours to the datasets caused mean error increases of up to 51.27%. Thus, this work contributes by comparing both types of thermography, analysing thermal aspects of the temperature data that influences the neural network’s estimation process, and suggesting alternatives to improve its accuracy.