Detecting Mammographically-Occult Cancer in Women with Dense Breasts Using Digital Breast Tomosynthesis

Most women in the USA who have dense breasts at screening mammography receive a letter notifying them that mammography is less effective for them and having dense breasts increases the risk of breast cancer. The letter advises women to talk with their physician whether they should have additional screening with ultrasound or magnetic resonance imaging (MRI). The possible benefit of additional screening is detecting a mammographically occult (MO) cancer. However, the likelihood that a woman has a missed cancer is not known. Thus, women are left with a difficult decision, balancing the uncertain potential benefit of additional screening against the known costs. These known costs are financial (as some states do not cover the supplemental screen) and the risk of an unnecessary biopsy, as the specificity of ultrasound and MRI are lower than mammography. We have developed a novel technique using a Radon Cumulative Distribution Transform (RCDT) to detect MO cancers. The RCDT can highlight subtle suspicious signals by detecting asymmetries between the left and right mammograms. Our technique achieved an area under the ROC curve of 0.81 using screening mammograms. Digital breast tomosynthesis (DBT), a pseudo-3D imaging technique, is replacing mammography in the USA, because of its higher sensitivity and specificity. However, MO cancers still exist in DBT. The goal of our research is to develop imaging biomarkers for MO cancers on screening DBT of women with dense breasts. This would allow women to know the likelihood that they have an MO cancer and, thereby, allow them to make a more informed choice regarding supplemental screening. The key difference between DBT and standard 2D mammography is the available information in the z-direction. Such additional information provides advantages for cancer detection, but it also adds technical complexity when applying RCDT on DBT images. There are three ways to process DBT exams for RCDT: 1) applying RCDT on 2D DBT slices, 2) applying RCDT on synthetic mammograms from DBT, and 3) applying the 3D RCDT on DBT volumes. To develop imaging biomarkers for MO cancer in screening DBT, we need to investigate the optimal method to process DBT for RCDT. We will develop imaging biomarkers for the three methods using a developmental dataset of 900 MO cancer cases (clinical cases read as normal, but the woman has breast cancer detected on her next screening DBT) and 1800 cases (clinical cases read as normal and the woman does not have breast cancer detected on her next two screening DBTs). We will utilize a 2D convolutional neural network (CNN) and a 3D CNN as robust classifiers to analyze the RCDT processed DBT for MO cancer detection. Using a 5-fold cross-validation, we will train CNNs for each method and find the optimal method to process DBT for MO cancer detection. Finally, we will use an independent dataset of 100 cases to validate the classifier. If we are successful, then up to 15 million women each year who have dense breasts will have needed information upon which to base their decision for getting supplemental screening.