Evaluación de modelos de redes neuronales sobre imágenes termográficas para la detección del cáncer de mama

Abstract

Abstract: Breast cancer represents one of the leading public health challenges worldwide, being the most commonly diagnosed cancer among women with approximately 2.3 million new cases reported in 2020; however, access to early detection methods such as mammography remains limited in many regions due to economic and geographic barriers. This research evaluates convolutional neural network models on thermographic images for the early detection of breast cancer, using the public DMR-IR dataset (272 images). The methodology combines transfer learning techniques through feature extraction with pretrained architectures (EfficientNet-B0 and ResNet50) and Support Vector Machine (SVM) classifiers. The ResNet50 + SVM model achieved an AUC-ROC of 0.9894, indicating a discriminative capacity very close to that of a perfect classifier, along with a precision of 92.31% and a specificity of 96.30% and an accuracy of 92.68%. Additionally, a prototype with a graphical user interface was developed, enabling the practical implementation of the system as a computer-aided diagnosis tool.