Tools based on deep learning models have been created in recent years to aid radiologists in the diagnosis of breast cancer from mammograms. However, the datasets used to train these models may suffer from class imbalance, i.e., there are often fewer malignant samples than benign or healthy cases, which can bias the model towards the healthy class. In this study, we systematically evaluate several popular techniques to deal with this class imbalance, namely, class weighting, over-sampling, and under-sampling, as well as a synthetic lesion generation approach to increase the number of malignant samples. These techniques are applied when training on three diverse Full-Field Digital Mammography datasets, and tested on in-distribution and out-of-distribution samples. The experiments show that a greater imbalance is associated with a greater bias towards the majority class, which can be counteracted by any of the standard class imbalance techniques. On the other hand, these methods provide no benefit to model performance with respect to Area Under the Curve of the Recall Operating Characteristic (AUC-ROC), and indeed under-sampling leads to a reduction of 0.066 in AUC in the case of a 19:1 benign to malignant imbalance. Our synthetic lesion methodology leads to better performance in most cases, with increases of up to 0.07 in AUC on out-of-distribution test sets over the next best experiment.
8 January 2021 in IEEE Transactions on Medical Imaging
Early breast cancer screening through mammography produces every year millions of images worldwide. Despite the volume of the data generated, these images are not systematically associated with standardized labels. Current protocols encourage giving a malignancy probability to each studied breast but do not require the explicit and burdensome annotation of the affected regions. In this work, we address the problem of abnormality detection in the context of such weakly annotated datasets. We combine domain knowledge about the pathology and clinically available image-wise labels to propose a mixed self- and weakly supervised learning framework for abnormalities reconstruction. We also introduce an auxiliary classification task based on the reconstructed regions to improve explainability. We work with high-resolution imaging that enables our network to capture different findings, including masses, micro-calcifications, distortions, and asymmetries, unlike most state-of-the-art works that mainly focus on masses. We use the popular INBreast dataset as well as our private multi-manufacturer dataset for validation and we challenge our method in segmentation, detection, and classification versus multiple state-of-the-art methods. Our results include image-wise AUC up to 0.86, overall region detection true positives rate of 0.93, and the pixel-wise ${F}_{{1}}$ score of 64% on malignant masses.