Aplicação de redes neurais convolucionais a imagens de modelos térmicos para a estimativa de tamanho e posição de inclusões que simulam tumores mamários

Abstract

Breast cancer is the most prevalent and deadly neoplasm among women worldwide, highlighting the urgent need for early detection strategies that are accurate, safe, and accessible. Mammography, while the gold standard, has limitations, including repeated radiation exposure, reduced sensitivity in dense breasts, and reliance on expert interpretation, while ultrasound and magnetic resonance imaging increase cost and complexity. Infrared thermography emerges as a non-invasive, low-cost, and radiation-free alternative, but visual inspection alone lacks the sensitivity to detect deep or small tumors. In this context, machine learning methods have shown promise in extracting subtle thermal patterns beyond human perception. This study implements and trains convolutional neural networks in Python using the Tensorflow library to learn the position and diameter of a spherical inclusion in thermal models, using steady-state temperature maps generated through finite element simulations in Ansys Mechanical. The methodology is applied to two thermal models: i) parallelepiped: simulating an experiment performed on a silicon parallelepiped with inclusions of spheres heated by thermal resistors, the neural network trained with 270 matrices (27x39) achieved an accuracy of 99.10% and a sensitivity of 98.10%, and was able to correctly predict the position and diameter of the inclusion when applied to an experimental thermogram. ii) anatomical model: simulating metabolically active tumor tissue immersed in anatomical breast geometry representing healthy breast tissue, the neural network trained with 965 matrices (50x50) achieved an accuracy of 97.63% and a sensitivity of 95.62%. These results demonstrate that a neural network can reliably learn geometric and physical characteristics related to inclusions (which simulate the presence of tumors) from low-resolution thermal data, indicating the feasibility of using artificial intelligence applied to thermal images for tumor detection.