comparação de modelos de CNN na detecção de câncer de mama utilizando imagens infravermelhas multiângulo

Abstract

Breast cancer is the second most common type among women, representing a major challenge for the healthcare field. Although mammography is widely used as the primary diagnostic method, its reliability is limited in younger patients. In this context, this study investigates the use of machine learning algorithms applied to thermographic images as a complementary alternative for early diagnosis. Using the DMR-IR database and pre-trained Convolutional Neural Networks (CNNs), architectural modifications found in the literature were applied to evaluate them in a multi-input scenario with different breast angles. Among the models tested, DenseNet achieved the best performance, with 94.44% accuracy, an F1-score of 0.9474, and 100% sensitivity, demonstrating its ability to correctly identify all positive cases. The results indicate that increasing neuron density can positively affect performance, provided the architecture can handle the greater information volume without overfitting — as seen with DenseNet, in contrast to EfficientNet, which showed limitations in this regard. Furthermore, the multi-input approach proved effective by enabling the analysis of regions previously inaccessible with single-input architectures. Therefore, the use of machine learning with thermographic images appears promising as an auxiliary tool in the early detection of breast cancer.