New research suggests that an emerging artificial intelligence (AI) software for mammography may lead to significantly enhanced detection for breast cancer and significantly reduced screen-reading workload for radiologists.
For the randomized, multicenter controlled trial, recently published in Lancet Digital Health, researchers compared the use of adjunctive AI screening (Transpara version 1.7.0, ScreenPoint Medical) in 53,043 women versus standard double reading in 52,872 women. The entire cohort (median age of 53.7) was drawn from four screening facilities in southwest Sweden, according to the study.
The researchers found that use of adjunctive AI facilitated a 29 percent increase in breast cancer detection (6.4 per 1,000 participants) in comparison to unassisted double reading by radiologists (5 per 1,000 participants).
In contrast to double reading by radiologists, the study authors pointed out that adjunctive AI also led to a 24 percent higher detection of invasive breast cancers, including 58 more T1 cancers and 46 more cases of lymph-node negative breast cancer.
“The large increase in detected small, lymph-node negative, invasive cancers suggests that downstaging by earlier detection with use of AI is possible, which could be of clinical benefit since stage has a major influence on breast cancer treatment and prognosis,” wrote lead study author Veronica Hernstrom, M.D., who is affiliated with the Radiology Department at Skane University Hospital in Malmo, Sweden, and the Diagnostic Radiology Department at Lund University in Lund, Sweden, and colleagues.
Adjunctive AI use was associated with a slightly higher recall rate (2.1 percent vs. 1.9 percent) but there were only seven more cases of false positives with adjunctive AI in contrast to unassisted double reading (772 vs. 765), according to the study authors. The researchers also noted a 5 percent higher positive predictive value (PPV) for AI recall cases (30.5 percent vs. 25.5 percent).
Three Key Takeaways
1. Increased cancer detection. Adjunctive AI screening led to a 29 percent increase in breast cancer detection compared to unassisted double reading by radiologists, particularly improving detection of small, lymph-node negative, invasive cancers that may allow for earlier intervention.
2. Reduced radiologist workload. AI-assisted screening resulted in a 44 percent reduction in screen-reading workload, potentially allowing radiologists to focus more on complex cases and patient-centered tasks.
3. Slightly higher recall rate but improved PPV on recalls. AI use was associated with a slightly higher recall rate (2.1 percent vs. 1.9 percent), but also a 5 percent higher positive predictive value (PPV) for AI recall cases, suggesting better efficiency in identifying true-positive cases.
Additionally, with the use of adjunctive AI to triage cases, the study authors found a 44 percent decrease in screen readings in comparison to the unassisted double reading of mammograms (61,248 vs. 109,692).
“The large reduction in screen-reading workload made possible by the AI-supported screen-reading procedure would free up time for breast radiologists to spend on more complex patient-centered tasks,” emphasized Hernstrom and colleagues.
(Editor’s note: For related content, see “Can Mammography-Based AI Enhance Breast MRI Use in Patients with Intermediate Risk for Breast Cancer?,” “Can AI Bolster Breast Cancer Detection in DBT Screening?” and “Mammography Study Shows Merits of AI for Improving Breast Cancer Detection and Effectiveness of Recalls.”)
In regard to study limitations, the authors conceded the use of single mammography and AI vendors. They also cautioned about extrapolation of the study findings to broader populations, noting the low baseline recall rates for mammography screening in Sweden.
