Breast cancer is the most lethal neoplasm among women in Brazil. Therefore, early detection is essential to initiate treatment promptly and reduce the risk of death. A public database of breast thermographic images was used for a classification task to distinguish between healthy and sick individuals. This process was conducted by a model based on transfer learning, in which a convolutional neural network pre-trained in a big dataset (ImageNet) is employed as a feature extractor and an artificial intelligence algorithm is trained on its top. To verify whether there are differences between different artificial intelligence architectures and approaches, two different convolutional neural network architectures were tested as feature extractors (VGG16 and ResNet), with two different approaches of artificial intelligence algorithms: the approach by neural networks and the approach by classical machine learning algorithms(Random Forest, Gradient Boosting, Support Vector Machine); with 8 models trained in total. In order to obtain statistically significant results, all models were trained in 5 different folds on cross-validation with 5 different seeds, yielding 25 distinct results for each model. The results indicated that the model that had the best performance in general was the approach that uses ResNet50 as an architecture with neural networks on top. Paired statistical tests, non-paired tests, and graphical analysis indicated that the models using ResNet50 as a feature extractor were superior to VGG16. Additionally, themodels that used neural networks on top of the classifier had an advantage compared to those that used classical machine learning algorithms.