Classification of Breast Lesions Using Mammary Sinograms and Deep Learning

Breast cancer is a disease that affects many women worldwide. Therefore, early detection of breast lesions is essential for effective diagnosis and treatment. In this context, digital breast tomosynthesis (DBT) is a promising technique for improving lesion detection compared to conventional mammography. However, image reconstruction can lead to a loss of information. Sinograms provide sufficient information for identifying and localizing lesions while reducing the noise and artifacts associated with the image reconstruction process. In this paper, we propose to employ image processing and deep learning techniques, especially convolutional neural networks (CNN), to identify patterns and extract relevant information from DBT sinograms. The objective is to facilitate diagnosis and treatment through the analysis of sinograms, allowing the identification of patterns and the extraction of useful information for continuous monitoring and evaluation of treatment over time. The results show that the ResNet50 and ResNet18 models have demonstrated a solid performance in classifying breast lesions from sinograms, with outstanding results in accuracy, recall, and F1-score of 94.96%, supporting their efficacy in this task. It is essential to highlight that although there are studies on sinograms in other fields, their application in diagnosing breast lesions has not yet been explored.