Since digital breast tomosynthesis (DBT) was approved for clinical use by the FDA in 2011 and Medicare coverage began in 2015, radiology practices and screening programs in the United States have rapidly adopted the technology. The most recent MQSA statistics suggest DBT-capable machines account for roughly 95% of installations in certified facilities nationally. Compared with full field digital mammography (FFDM), DBT maximizes the identification of biologically significant disease while mitigating overdiagnosis. However, tradeoffs exist, especially for interpretation times in already-stretched breast imaging workloads. Dr. Ben Pettus, MD, PhD, a breast imaging specialist and fellowship-trained radiologist, sums up the dilemma: "We've gone from four images, maybe six images on a case, to hundreds of images. It raises the risk for fatigue or distraction."
Dr. Pettus, who also holds a PhD in Molecular and Cellular Biopathology, leads breast imaging for a mixed academic and private practice group in coastal Virginia, Peninsula Radiological Associates. They operate as a multi-vendor (GE, Siemens, and Hologic) and multi-site organization spanning a wide rural-to-urban network. As his practice successfully adopted DBT, Dr. Pettus continued to stay current on the latest tools and tech which led him to exploring artificial intelligence to solve one key clinical bottleneck: "I've always kind of respected the European model of having a secondary read, having two radiologists look at something," he said. "That's never been really possible in the US before. To me, this was a patient safety issue." An AI system that promises to review every slice on every case, every time, would afford his team a second set of eyes.
After roughly four years of running Transpara on every screening mammogram in conjunction with DBT adoption, the practice’s outcomes tell a promising story: breast cancer detection rates have nearly doubled, average tumor size at detection has been cut in half, reading efficiency is up an estimated 20 to 30 percent, and recall rates have fallen.
As with any new technology, confidence builds with use. "It takes a while to just know what [Transpara is] going to call, what you're going to call. After a while, it's like having a really high-end fellow working with you; somebody who's a very well-trained radiologist who's got a good eye for the morphology, and they went through everything as well." For private-practice radiologists without residents or fellows asking "What's that?", that second perspective is something they've never had. Four years in, across the whole group:
Breast cancer detection has nearly doubled. "In the 2D environment, it was about 3 to 4 out of a thousand. It's climbed even further with adding the AI to the 3D; it’s roughly 7 to 8 out of a thousand now." Importantly, the gains aren't concentrated in a single breast imager: "I see the whole group, everybody's rates are going up."
Cancers are found smaller and earlier. Average mass size at detection has dropped from roughly 12 to 14 mm in the 2D era to about 7 mm, with finds as small as 3 mm. "It's kind of amazing, because it's even smaller than you could find on an MRI."
Recall rates went down, not up. The group's callback rate has moved from roughly 8 percent to closer to 6 percent, comfortably on the low end of the typical US range, even while detecting more cancers.
Efficiency improved 20 to 30 percent. That surprised him most, because he deliberately did not deploy the AI for triage. The gains came from two places: less compulsive re-reading of clean negatives ("having the knowledge that this AI has gone through every slice as well... I found myself not obsessing as much to reread things"), and faster problem-solving on complex, lesion-everywhere cases, where the AI's concordant flags across CC and MLO views narrow the search. He also called it "phenomenal at calcifications in particular."
The cumulative effect on capacity is striking: "I probably get about three times as much work as I did 15 years ago. I actually have more things to look at, but I can get through more, faster. I'm still not having interval cancers or misses." And there's a human dividend: "less fatigue at the end of the day."
His bottom line: "The net effect to me is I'm able to find things a lot smaller, earlier, get more work done, and feel like I've added more value to the patients, because we're finding more things early. That's the whole goal. It's kind of a no-brainer. Just go for it, and like anything, just be careful and learn as you go."
To hear more from Dr. Pettus on AI adoption, as well as how he vetted the algorithms with his own dataset, watch the interview below:
Quotes have been lightly edited for clarity and length. Performance figures reflect Dr. Pettus's reported experience at his practice and are not generalized claims.