Despite a recent report calling into question the benefits of digital breast tomosynthesis (DBT) – sometimes called 3D mammography – as a screening tool for the early detection of breast cancer, the technology has already been widely adopted. In fact, there were 8,235 accredited DBT systems installed in the USA as of 1 November 2019. While DBT may provide benefits for patients – such as improved detection and fewer recalls – it does pose a major challenge for radiologists: an exponential increase in the number of images that must be read compared to 2D mammography (click on DBT image below left, © Siemens and Malmö University Hospital).

In a recent article published in Clinical Radiology, a group from Cambridge University has reviewed the potential of artificial intelligence (AI) to reduce the load on the radiologist. They introduced the topic of machine learning and, in particular, the application of artificial neural networks (see above right, © The Royal College of Radiologists). Such a network consists of consecutive layers, and comprises an input layer, multiple hidden layers and an output layer. Training of the network requires vast amounts of input data – such as DBT images – for successful clinical implementation.

With the upcoming annual meeting of the Radiological Society of North America (RSNA) in a week’s time, companies are gearing up to demonstrate their AI solutions for analyzing DBT images. In an opinion piece for AuntMinnie.com, Randy Hicks has described the benefits of ProFound AI for DBT from iCAD which received FDA clearance in 2018. He cited a recent article in Radiology: Artificial Intelligence by Emily Conant in which ProFound improved sensitivity by 8% and reduced radiologist reading time by 57% for women with dense breasts.

Writing in Radiology, Krzysztof Geras illustrated how a commercially available system can use AI to enhance synthetic images generated from the DBT volume. Below right are four images (© RSNA), where AI has been applied to enhance the visibility of invasive ductal carcinomas (see arrows).

In a litigious climate such as the USA, the widespread adoption of DBT+AI could potentially complicate the question of liability. As Linda Moy has asked: “Where does the responsibility lie if there is a misdiagnosis? Who is liable? Is it the radiologist or the vendor who developed the software?” Despite the ethical, social and legal implications of applying AI to DBT images, Michael Fuchsjäger is optimistic, suggesting: “AI algorithms should be expanded to cover other fields of breast imaging with huge amounts of data such as MRI and 3D automated breast ultrasound (ABUS).” Time will tell how this story unfolds.