Women’s Imaging: A Different Type of House Call
By Beth W. Orenstein
Radiology Today
Vol. 23 No. 6 P. 28

Can AI researchers improve breast imaging by visiting data at home?

Interest is booming in using AI to assist with the detection and diagnosis of breast cancer, which, like many cancers, is most successfully treated when caught early. One reason for AI’s growing demand is that today’s breast imaging technologies, such as 3D digital breast tomosynthesis, produce significantly larger volumes of imaging data per patient than 2D mammography. Why wouldn’t breast imagers welcome AI tools that help them sort through this large volume more easily and go beyond traditional computer-aided detection, which was introduced in the late 1990s?

A study published in February 2020 in the journal The Breast found that AI is helping meet the demand for more accurate detection, classification, and prediction of how breast tumors will behave. But, while AI has made some inroads in breast imaging, there is a problem: a lack of high-quality data from racially, geographically, and socioeconomically diverse populations.

Research published in 2020 in the Journal of the American Medical Association found that cohorts from California, Massachusetts, and New York are disproportionately represented in training sets for deep learning algorithms, with next to no representation from other states. Large portions of the United States are “data deserts,” meaning their populations have little representation in the medical AI literature, says Aaron Mintz, MD, an assistant professor of radiology at Washington University School of Medicine in St. Louis and medical director at Flywheel, which offers comprehensive data management solutions for researchers.

If these populations are not well represented in the data used to develop breast imaging AI, the performance and usefulness of the technology will suffer, Mintz says. “Making sure that different populations are well-represented—being sure you are including those with medical histories, surgical histories, patients who have implanted devices such as pacemakers, and other anatomical variations—is crucial to the success of any AI tool used in breast imaging,” he says.

Learning to Share
One of the barriers to building robust, diverse data sets is sharing data in a way that complies with regulations and respects patient privacy. Given this need, providers are resistant to allowing their imaging data to leave the security of their own PACS, Mintz says, making it difficult for researchers to access the diverse data they need to build AI models without bias.

Most machine learning techniques require local data sets to be uploaded to a single server. Mintz believes there is a solution to this problem: federated learning. Federated learning makes it possible for researchers to access data where it resides. Providers do not have to share sensitive patient data across institutions.

“The data never leaves,” Mintz explains. “The data never goes anywhere.” The concept, he says, opens a world of possibilities for collaborative research on biomedical data.

The need for training AI on diverse populations is urgent. Increasing demands for breast imaging coupled with a growing shortage of breast imagers present an opportunity for breast imaging AI to ensure access to high-quality breast cancer screening and detection for all patients, Mintz says. But this opportunity will not be realized if all populations who stand to benefit from the technology are not represented in its development.

“To cite just one example, it is well established that mortality rates from breast cancer are higher for Black and African women, despite comparable screening rates to white patients,” Mintz says. “The reasons for this are multifactorial and include, for example, differences in breast density and how patterns of breast density may contribute to cancer risk.”

While current assessments of breast density are qualitative, AI can evaluate breast tissue quantitatively across numerous measures that will likely be useful in assessing cancer risk. To realize this potential benefit for Black women, it will be crucial to make sure that the population is well represented in training and validation data for the technology. If that is accomplished, “improving breast AI will improve overall breast imaging quality,” Mintz says.

Built-In Safeguards
With federated learning, “developers can visit the data, do an analysis or some training, and then take the results of that and bring that back without ever inspecting the images themselves,” Mintz says. Institutions may be rather reluctant to package up 50,000 exams and send them off somewhere to be included in AI training, but he believes they could be more open to someone coming to them and examining those same 50,000 images for training.

As part of the process, safeguards are built into the programs so that institutions do not have to worry about their data being compromised or affected in any way by the visit from the AI developers. So far, Mintz says, Flywheel has found that institutions are less reluctant to let developers work with data on their own infrastructure behind their firewalls.

“Federated learning dramatically reduces the risk of unauthorized access to patient data and the risk of patient data leaving the institution and being used in a noncompliant way,” he says.

A federated approach does necessitate that data are prepared and organized to standards so that they are interoperable and to mitigate irrelevant differences across data sets that do not impact AI performance. This requires a multimodal data science infrastructure that supports not only data ingestion and preparation but also distributed computation. With the appropriate technology stack, Mintz says, “the data can be curated and cleaned and prepared to a threshold that the AI researchers or developers expect. Then you can just put the algorithm to work. It comes and does its training.”

According to Mintz, federated learning can be applied to nearly all imaging, but Flywheel’s current development work is emphasizing breast mammography for several reasons: one is that breast imaging has always been a leader in radiology practices, he says. “Breast imaging is more standardized than any other imaging modality,” Mintz says. Second, anatomically, “there’s less complexity.” Another reason to start with breast imaging, he notes, is the aforementioned “substantial disparities across patient populations in terms of outcomes and access.”

Mintz says Flywheel has been working on a platform that facilitates the use of federated learning for building breast imaging AI tools for more than a year and plans to showcase a demonstration of the concept at the RSNA annual meeting in November.

— Beth W. Orenstein of Northampton, Pennsylvania, is a freelance medical writer and regular contributor to Radiology Today.