Commercial Artificial Intelligence (AI) Tool for Screening Digital Breast Tomosynthesis: Factors Associated With AI-Based Breast Cancer Detection.
Authors
Affiliations (3)
Affiliations (3)
- Department of Radiology, Massachusetts General Hospital, 55 Fruit St, WAC 240, Boston, MA 02114.
- Present affiliation: Department of Biomedical Science, Seoul National University, Seoul, South Korea.
- Present affiliation: Department of Molecular Biosciences and Bioengineering, University of Hawaii at Manoa, Honolulu, HI.
Abstract
<b>BACKGROUND.</b> Research on artificial intelligence (AI)-based computer-assisted detection and diagnosis (CADe/CADx) algorithms has focused primarily on digital mammography rather than on digital breast tomosynthesis (DBT). Additionally, DBT-related studies have not comprehensively stratified performance by cancer characteristics. <b>OBJECTIVE.</b> This study's purpose was to evaluate factors associated with cancer detection for a commercial AI-based CADe/CADx algorithm for DBT interpretation. <b>METHODS.</b> This retrospective study included consecutive screening DBT examinations performed from January 2016 to June 2019 that were classified in the institution's breast imaging reporting system as true-positive (<i>n</i> = 500), false-negative (<i>n</i> = 100), true-negative (<i>n</i> = 4400), or false-positive (<i>n</i> = 2500) cases based on radiologists' clinical interpretations and 1-year follow-up outcomes. A commercial AI-based CADe/CADx DBT algorithm (Genius AI Detection 2.0 software, Hologic) analyzed examinations for investigational purposes. A breast imaging radiologist reviewed radiologist true-positive and false-negative examinations with positive AI results to determine whether AI-annotated lesions corresponded to the locations of diagnosed cancers. Factors associated with AI detection were evaluated. <b>RESULTS.</b> The study included 7500 patients (mean age, 59 ± 12 [SD] years) who underwent 7500 DBT examinations. AI detected and correctly localized cancers in 89.8% (449/500) of radiologist true-positives and 32.0% (32/100) of radiologist false-negatives. AI correctly categorized 55.1% (2426/4400) of radiologist true-negatives and 38.9% (972/2500) of radiologist false-positives as negative. Among radiologist true-positives, AI detected and correctly localized 92.4%, 81.6%, 86.7%, and 85.7% of invasive ductal carcinomas, invasive lobular carcinomas, other invasive carcinomas, and cases of ductal carcinoma in situ, respectively (<i>p</i> = .049); AI detected and correctly localized 84.4%, 91.5%, and 95.2% of grade 1, 2, and 3 invasive carcinomas, respectively (<i>p</i> = .03). Among radiologist false-negatives, AI detected and correctly localized 41.2% and 8.3% of cancers with versus without a mammographic finding reported during later diagnostic workup (<i>p</i> = .04). AI detection showed no significant association with age, race, breast density, mammographic finding type, tumor size at surgery, hormone receptor status, or lymph node involvement in either group (<i>p</i> > .05). <b>CONCLUSION.</b> AI detected and correctly localized 89.8% of radiologist true-positive and 32.0% of radiologist false-negative cases. Certain cancer characteristics were associated with AI detection. <b>CLINICAL IMPACT.</b> The results may help radiologists understand the algorithm's strengths and limitations and inform algorithm refinement efforts.