van der Schaar Lab

Machine learning for mammography article named “Best of 2020” by JACR

The van der Schaar Lab is delighted to announce that an article authored by Ph.D. student Trent Kyono and Prof. Mihaela van der Schaar, in collaboration with Prof. Fiona Gilbert from the University of Cambridge’s Department of Radiology, has received a “Best of 2020” award from the Journal of the American College of Radiology (editor-in-chief: Ruth C. Carlos, MD, MS).

Trent Kyono
Prof. Mihaela van der Schaar
Prof. Fiona Gilbert

First published in the journal’s January 2020 issue, the article, entitled “Improving Workflow Efficiency for Mammography Using Machine Learning,” was chosen by a committee of members of the JACR Editorial Board as the best article in the journal’s Data Science category during this past year. The board annually recognizes five articles, one for each of the journal’s areas of interest; selection criteria include importance to the specialty and lucidity of presentation.

The article presents a study conducted with the aim of determining whether machine learning could reduce the number of mammograms radiologists must read by correctly identifying normal mammograms and only referring uncertain and abnormal examinations for radiological interpretation.

Using machine learning to empower (rather than replace) radiologists

Breast cancer is the most prevalent cancer diagnosed in women, and the number of women requiring screening and diagnostic mammography continues to increase.

A key priority in cancer diagnosis is managing the workload of radiologists to optimize accuracy, efficiency, and costs. As examination volumes and time of interpretation increase with newer screening technologies such as digital breast tomosynthesis, radiologists will be forced to read more patients in less time. It is increasingly difficult to ensure that radiologists can devote the right amount of time to reading scans that actually need their attention, rather than wasting time reading scans that do not require their expertise—especially since doing so can hinder their overall performance.

Meanwhile, a wide variety of AI and machine learning tools for detection and localization of lesions in mammograms have been developed in recent years, with early computer-aided detection giving approaches way to more advanced and powerful technologies driven by deep learning, such as convolutional neural networks. Most of these tools, however, are designed with one aim in mind: detecting and diagnosing cancer effectively as early as possible. This aim bears an implicit assumption that such technologies can and will outperform (and, therefore, replace) expert human radiologists.

In reality, however, such tools still require manual examination and validation by an expert radiologist, and bear additional software interfacing requirements for radiologists, significantly increasing the average reading time per patient. Mechanisms that can effectively reduce the number of examinations to be read by radiologists have not been provided.

By contrast, the method introduced in the lab’s award-winning 2020 JACR article introduces a hybrid approach that inherently accommodates the benefits of (and need for) a sort of “division of labor” between machine learning/AI and human radiologists, with both playing to their respective strengths.

The system in question, Autonomous Radiologist Assistant (AURA), is a machine learning method capable of triaging a subset of examinations as negative with extremely high accuracy and referring the rest for expert consultation—thereby significantly reducing the daily interpretive workload of radiologists and freeing up their time to focus on more suspicious examinations and diagnostic work-ups.

By combining the use of a convolutional neural network with multitask learning, the AURA system is able to extract imaging features from mammograms that mimic the radiological assessment provided by a radiologist, the patient’s nonimaging features, and pathology outcomes. A deep neural network is then used to concatenate and fuse multiple mammogram views to predict both a diagnosis and a recommendation of whether or not additional radiological assessment is needed.

A system-level illustration of AURA.

In Step 1, all four of a patient’s mammograms are passed to a multi-task feature extractor (a convolutional neural network) trained using multi-task learning to emulate the mammographic predictions of a radiologist. In Step 2, the predictions extracted from the individual mammograms are fused using a deep neural network to provide a multi-view diagnosis on the basis of both a patient’s imaging and non-imaging features. Lastly, AURA considers the radiological predicted features (multi-task outputs from Step 1), the multi-view cancer prediction (from Step 2), and the patient’s non-imaging features to issue a recommendation for patient triage.

The authors presented the AURA system on a private mammography dataset, and demonstrated that AURA could autonomously triage patients between a machine learning classifier and a radiologist to effectively reduce the number of negative mammograms that a radiologist reads while maintaining an admissable negative predictive value (above 99%). They demonstrated AURA’s ability to filter patients from the radiologist with attributes known to be associated with lower likelihoods of cancer, such as younger age and lower breast density. Crucially, they established through validation that AURA was capable of significantly decreasing the workload for the radiologist by 34% in a diagnostic setting, and by nearly 91% for a screening-like setting.

In summary, through methods such as AURA, machine learning could be successfully leveraged to reduce the number of normal mammograms that radiologists need to read without degrading diagnostic accuracy. This was the first study investigating machine learning methods for reducing the number of negative mammograms a radiologist needs to read without replacing the radiologist entirely.

Improving Workflow Efficiency for Mammography Using Machine Learning

Trent Kyono, Fiona J. Gilbert, Mihaela van der Schaar

Journal of the American College of Radiology, 2020

Further reading

If you’d like to learn more about our lab’s work related to cancer, feel free to browse the resources below, which include 2 engagement sessions with clinicians and 2 written overviews introducing a range of projects.

If you are a clinician and would like to learn more about how machine learning can be applied to real-world healthcare problems, please sign up for our Revolutionizing Healthcare online engagement sessions (no machine learning knowledge required).

For a full list of the van der Schaar Lab’s publications, click here.

Nick Maxfield

From 2020 to 2022, Nick oversaw the van der Schaar Lab’s communications, including media relations, content creation, and maintenance of the lab’s online presence.