Application of Deep Learning in Breast Cancer Imaging

With 685.000 deaths, breast cancer is the leading cause of cancer mortality worldwide among women in 2020. Breast cancer imaging plays a major role in reducing this excessive number of deaths. Screening programs have been set up to detect breast cancer early, enabling easier treatment and higher survival rates than cancers detected at later stage. Moreover, breast cancer imaging techniques are essential for monitoring and evaluating cancer treatment.

Since the breakthrough of deep convolutional neural networks (CNNs) in 2012, deep learning (DL) techniques and, in particular, CNN's were rapidly implemented in breast radiology to be used as detection/diagnosis (CAD) systems, hopefully surpassing the radiologist performance level. In recent years, although many applications and submitted papers are still optimizing these CAD systems, the role of AI and DL has exceeded the early objective of CAD systems to aid radiologists only in these two tasks. Nowadays, DL is used in a plethora of different tasks, such as image reconstruction and generation, cancer risk prediction and prediction, and assessment of therapy response.

In this review article, we give an overview of the current role of DL in these various tasks for the four main imaging modalities for breast cancer, namely mammography, digital breast tomosynthesis (DBT), ultrasound, and magnetic resonance imaging (MRI). Although the role of DL in these different imaging modalities is comparable, the difference in the size of the datasets and the level of validation and evaluation of the studies differ much. Whereas mammography and DBT studies nowadays have large datasets in the order of (tens of) thousands of patients from different medical centers, MRI studies only rarely exceed 500 patients and are often from a single center. The size of ultrasound datasets is in between the other modalities' dataset sizes, in the order of some thousands of patients. This favors the AI performance in mammography and DBT studies, as generally larger datasets and data from different centers, lead to better performing and better generalizing DL models. The same observation can be made for the level of validation and evaluation per modality.

If you want to know more details about all the roles of DL in breast cancer imaging across the different imaging modalities, and the future role of these techniques in clinical practice, we strongly encourage you to read the paper! In the meantime, we will continue to do research into DL techniques for breast cancer imaging, especially MRI, to detect cancer as early as possible and consequently reduce the yearly number of breast cancer deaths.