Invited Commentary: The Power and Promise of Artificial Intelligence for Digital Breast Tomosynthesis

See also the article by Goldberg et al in this issue.

Digital breast tomosynthesis (DBT), which was approved for use in the clinical setting by the U.S. Food and Drug Administration (FDA) in 2011, is currently being used in more than 80% of breast imaging facilities in the United States (1). With DBT, multiple low-dose x-rays are obtained along an arc and reconstructed into a stack of thin sections of the breast (2). DBT images are interpreted in conjunction with either full-field digital mammography (DM) images, which are acquired separately, or synthetic two-dimensional (2D) mammographic images, which are reconstructed from the DBT images, sparing the patient the radiation dose from a separate 2D acquisition. Multiple studies have shown that DBT in conjunction with 2D mammography leads to higher cancer detection rates and lower false-positive rates compared with 2D mammography alone (2,3). However, the addition of DBT to DM has led to increased interpretation times (4). In addition, there are concerns about the quality of synthetic 2D mammography compared with the quality of DM (5).

In their comprehensive review article “New Horizons: Artificial Intelligence for Digital Breast Tomosynthesis” published in this issue of RadioGraphics, Goldberg et al (6) discuss how artificial intelligence (AI) has the potential to heighten the advantages of DBT while also mitigating its limitations. AI tools can help radiologists increase their sensitivity and specificity for the detection of breast cancer with DBT, create synthetic 2D mammographic images that accentuate suspicious findings, and reduce the radiation dose of DBT, among other applications. Importantly, AI tools can increase efficiency by decreasing interpretation time and/or triaging out normal mammographic studies from the radiologist’s queue.

At present, there are more than 10 FDA-cleared applications for DBT, which are intended for breast density assessment, lesion detection and diagnosis, and triage (7,8). The breast density applications provide quantitative assessments of breast density in addition to categorization according to the Breast Imaging and Reporting Data System (BI-RADS) atlas. The lesion detection and diagnosis applications are used to identify suspicious findings with DBT and provide likelihood of malignancy scores at the finding, breast, and/or case levels. The one triage application for DBT is used to flag mammographic studies that have at least one suspicious finding. This information can be used to generate customized work lists for radiologists and/or prioritize certain imaging studies.

What does the future hold for AI and DBT? As Goldberg and colleagues (6) discuss, many of the AI applications, including those that are cleared by the FDA, have not yet been evaluated in real-world clinical settings; thus, their impact on patient care and patient outcomes remains unknown. We must ensure that existing AI tools are effective across diverse patient populations and different vendors. Furthermore, Goldberg et al (6) describe multiple other AI tools for DBT that are in the research and development phase—for example, those that predict long-term breast cancer risk on the basis of DBT imaging findings and those that remove normal breast tissue on images to improve breast cancer detection.

The review article by Goldberg et al (6) focuses on the use of AI for DBT, which is of particular interest given the high utilization of DBT for screening mammography. FDA-cleared applications are also available for DM (7,8). As Goldberg et al (6) discuss in their article, the applications for DM are less technically challenging to develop than those for DBT for several reasons: DM datasets require less computational capacity than DBT datasets, DM annotations are less complicated than DBT annotations, training datasets for DM are larger than those that are available for DBT, and there is less variability in the appearance of breast tissue across DM vendors than across DBT vendors.

FDA-cleared applications are also available for lesion detection and diagnosis with breast US and lesion diagnosis with breast MRI (7,9). Multiple other AI applications for US and MRI are in the research and development phase and include those used to predict axillary lymph node metastases in patients with biopsy-proven breast cancer, response to neoadjuvant therapy, and breast cancer recurrence and survival (9).

Given its unique focus on DBT, the review article by Goldberg and colleagues (6) is a welcome addition to the growing breast imaging literature on AI. The review provides a comprehensive and thorough discussion of AI applications for DBT and the current status of AI for DBT, highlighting the message that the power and promise of AI for DBT have yet to be realized. The FDA-cleared AI applications for DBT are intended for breast density assessment, lesion detection and diagnosis, and triage, but clinical evaluation of these tools is limited, and there are multiple other potential AI applications for DBT detailed in the Goldberg et al article (6). Given the breast imaging community’s experience with traditional computer-aided detection, which did not live up to expectations when it was deployed in real-world clinical settings, we remain cautiously optimistic about the potential of AI to improve quality and efficiency in breast imaging (10). Ultimately, the success of AI tools for DBT will depend on appropriate integration into clinical workflows and demonstration of added value to breast imaging practices.