A high specificity deep learning approach with focus on breast cancer screening
Pedro Vilares, João C. Ferreira, Luís A. Bastião Silva, Augusto Silva · 2021 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) · 2021
Breast cancer is the leading type of cancer in women and the second most common cancer overall. With thousands of breast screening exams being performed daily around the world, it is a time-consuming task for radiologists, that often find it difficult to analyze and classify them all, with most of exams turning out to be normal cases. The dedicated time to review negative cases could be applied to reviewing more complex cases that require additional care. In this work, the authors propose a support to diagnosis system, focused on mammography screening, based on deep Convolutional Neural Networks (CNNs) and an ensemble classifier, designed to relieve radiologists of normal cases. The architecture takes advantage of macroscopic full mammogram-level features, patch-level features, and patient metadata to output an image-level classification. The results translate the architectures’ high confidence on the exam being normal, relieving the radiologist from analyzing the exam, or suspicious, requiring a specialized radiologist’s attention for a final classification. The developed system was trained and validated using three public datasets (CBIS-DDSM, BCDR and INbreast) and achieved a final, exam-level AUC of 0.98, with a specificity of 96% and a sensitivity of 87%. A conclusion of this work is the possibility to reduce the radiologist’s workload, with potential to reduce the requirement of a second reader, creating further opportunities to review and analyze more complex cases.