OR WAIT null SECS
Combination also leads to a reduction in the false-negative rate.
Combining artificial intelligence (AI) with a single-view wide-angle digital breast tomosynthesis (DBT) image could improve breast cancer screening programs while reducing radiation exposure for women by roughly half.
In a new study published July 6 in Radiology, a team of investigators from Radboud University Medical Center in The Netherlands examined whether radiologists could effectively identify suspected breast cancers by interpreting single-view DBT instead of DBT plus digital or synthetic mammography.
Based on their analysis, they found radiologist performance did improve with the AI/single-view combination – and it could also save money.
“Using this DBT and AI setup could allow for a more cost-effective screening program with higher performance and lower workload for radiologists,” said the team led by Marta C. Pinto, MSc, a doctoral student in the department of radiology and nuclear medicine at Radboud. “Overall, our study adds to the growing body of evidence on the improvement of screening outcomes by adding an AI system to DBT for decision and navigation support, suggesting that use of single-view DBT combined with AI is a feasible DBT screening implementation option.”
For their study, the team created a case set of 190 DBT examinations from 4,750 DBT studies gathered from clinical DBT examinations conducted at Radboud between January 2016 and February 2018. The case set included 75 exams with malignant lesions, 25 with benign lesions, and 90 normal scans. The team also required a one-year normal follow-up or histopathologic assessment as a reference standard.
Fourteen radiologists interpreted the studies during two sessions separated by 4 weeks. The exams were a random mix of those being read with and without the use of AI that identified suspicious regions with calcifications and soft-tissue lesion and provided both region and whole-exam level of suspicious (LoS) scores.
According to their evaluations, using the AI tool improved the area under the curve for radiologists – the AI-aided AUC was 0.88 compared to unaided 0.85. In addition, incorporating the tool improved their sensitivity – 86 percent versus 81 percent, as well as reduced their false-negative rate by 27 percent.
However, their results did not point to any improvements in specificity. There was also not overall improvement in reading time, but Pinto’s team did discover that, when using the tool, radiologists spent 8 percent less time interpreting low-suspicion exams and 27 percent more on high-suspicion studies.
These findings, they said, point to the feasibility of using an AI-fueled single-view DBT to improve screening program performance. However, in an accompanying editorial, two experts from the University of Michigan said doing so should be approached with caution. They pointed out that the interpreting radiologists who improved with the AI tool initially had lower performance than the AI tool did alone, and the higher-performing radiologists exhibited a drop when using it.
“These results indicated that the study radiologists tended to be influenced by the AI [computer-aided detection] assessment and converged to the AI [computer-aided detection] assessment rather than surpassing it in this retrospective reader study,” said Heang-Ping Chan, Ph.D., director of the University of Michigan CAD-AI Research Laboratory, and Mark A. Helvie, M.D., Trygve O. Gabrielsen professor of radiology at Michigan Medicine. “This emphasizes the importance of rigorous validation of the generalizability of an AI [computer-aided detection] tool before clinical implementation, as well as periodic surveillance to ensure the consistency of its performance in the clinic over time.”
To overcome that hurdle, they said, radiologists should be trained to use the AI tool and complete a “test drive” to determine its performance in their clinical environment. Doing so would help them develop appropriate expectations for how the tool would perform and limit their over-reliance on it.
Ultimately, Pinto’s team, Chan, and Helvie agreed that additional research is needed to figure out how well these single-view DBTs will actually work in practice.
“We recommend that prospective, international, and multi-center screening studies be undertaken to determine that actual impact of the use of AI-aided single-view DBT for screening,” Pinto’s team concluded.
For more coverage based on industry expert insights and research, subscribe to the Diagnostic Imaging e-Newsletter here.