Emerging results from a new study evaluating artificial intelligence (AI) in digital breast tomosynthesis (DBT) screening for breast cancer shows double the positive predictive value (PPV) for cases with abnormal interpretation and over a 24 percent higher PPV rate for cases proceeding to biopsy.
For the retrospective study, recently published in Clinical Breast Cancer, researchers compared unassisted radiologist interpretation of 10,322 DBT screening mammograms vs. the use of adjunctive AI (ProFound AI V2.1, iCAD) in 6.407 DBT evaluations.
The study authors found that adjunctive AI use led to a lower abnormal interpretation rate (AIR) in comparison to unassisted DBT interpretation (6.5 percent vs. 8.2 percent). However, employment of adjunctive AI was associated with a significantly higher cancer detection rate (CDR) than radiologist assessment of DBT (6.1/1000 vs. 3.7/1000), according to the researchers.
“In our study, we show significant improvements in breast cancer detection, positive predictive values, specificity, and abnormal interpretation rates in a real-world clinical setting. Our findings align with previous works and confirm that AI support can be translated to the real-world with the potential to augment radiologist diagnostic accuracy in DBT screening programs,” wrote lead study author Joshua A. Nepute, M.D., an assistant professor of radiology and imaging sciences with the Department of Radiology and Imaging Sciences at the Indiana University School of Medicine in Indianapolis, Ind., and colleagues.
Adjunctive AI assessment of DBT offered a significantly higher PPV for AIR cases in contrast to unassisted radiologist interpretation (8.8 percent vs. 4.2 percent), according to the study findings. The researchers also noted a higher PPV for adjunctive AI when cases proceeded to biopsy (56.5 percent vs. 32.3 percent).
Three Key Takeaways
1. Improved cancer detection. The use of adjunctive AI in DBT screening nearly doubled the cancer detection rate compared to unassisted radiologist interpretation (6.1/1000 vs. 3.7/1000), including higher detection of invasive cancers (5.5/1000 vs. 3.3/1000).
2. Enhanced diagnostic accuracy. AI support significantly increased positive predictive value (PPV) for abnormal interpretations (8.8 percent vs. 4.2 percent) and biopsy-confirmed cases (56.5 percent vs. 32.3 percent), indicating more accurate identification of malignancies.
3. Reduced false positives. AI integration led to a lower abnormal interpretation rate (AIR) compared to radiologist-only review (6.5 percent vs. 8.2 percent), potentially reducing unnecessary follow-ups and over-diagnosis.
The study authors also noted that a subsequent subgroup analysis involving women with invasive cancers revealed a higher CDR for adjunctive AI (5.5/1000 vs. 3.3/1000)
“ … Although breast cancer screening programs have substantially reduced mortality, overdiagnosis of breast cancer remains an issue posing real harm to patients. In the subgroup analysis, we report increased detection of invasive cancers with AI suggesting AI support could play a role in reducing overdiagnosis rates,” posited Nepute and colleagues.
(Editor’s note: For related content, see “Mammography Study Compares False Positives Between AI and Radiologists in DBT Screening,” “What New Research Reveals About the Impact of AI and DBT Screening: An Interview with Manisha Bahl, MD” and “Mammography Study Suggests DBT-Based AI May Help Reduce Disparities with Breast Cancer Screening.”)
In regard to study limitations, the authors acknowledged that radiologists may have more time to read post-AI DBT studies with this part of the study period coinciding with restrictions on screening mammograms due to COVID-19. While the study data was derived from three rural institutions, the authors cautioned about extrapolation of the results to broader populations due to all three centers being associated with one institution. They also pointed out that White women comprised over 80 percent of one facility’s cohort and nearly 88 percent of patients from another site.
