Automatic Segmentation of Breast Thermal Images Using Mask R-CNN

Abstract

Clinical thermography has been widely proposed as a complementary tool for the early detection of breast cancer due to its non-invasive nature, low cost, and absence of ionizing radiation. However, one of the main technical challenges is accurately segmenting the breast region (BR), as the thermal similarity surrounding it can bias analysis. This study presents an automated approach for segmenting the BR in infrared thermal images using the Mask R-CNN neural network. The model was trained on a set of breast thermograms labeled with manually annotated masks. The resulting automatic segmentation enables accurate isolation of the region of interest (RoI), minimizing the influence of irrelevant areas and supporting a more objective thermal analysis. Evaluated using the DICE coefficient, the system achieved an average value of 90.98%, demonstrating its effectiveness in potential applications for computer-aided diagnosis systems and standardizing thermal image analysis for breast health.