Speaker
Description
Breast infrared thermography is a non-invasive imaging technique capable of capturing superficial temperature distributions associated with variations in metabolic and vascular activity, offering potential support for early breast cancer screening. However, variations in acquisition conditions, patient morphology, and temperature ranges across sessions introduce significant inconsistencies in the RGB color mapping of thermographic images, limiting their quantitative comparability and subsequent automated analysis. This study presents a comparative evaluation of three families of Deep Learning models: (i) Generative Adversarial Networks (GANs), (ii) Convolutional Neural Networks (CNNs), and (iii) Transformer-based architectures for RGB color normalization of breast thermographic images. The objective is to achieve a standardized color representation while preserving the underlying thermal information encoded in the temperature matrix. The proposed methods are quantitatively assessed using objective metrics related to color similarity, thermal reconstruction accuracy, and preservation of diagnostically relevant thermal patterns. Particular attention is given to minimizing color distortion without compromising spatial and thermal detail critical for infrared image interpretation. The results aim to identify the most suitable Deep Learning paradigm for color normalization in breast thermography and to provide practical recommendations for quantitative infrared imaging workflows, supporting downstream tasks such as segmentation, classification, and longitudinal thermal analysis in both clinical and research contexts.
