Table 2.
Summary of study characteristics for studies using AI for triage
| Study | Study design | Population | Mammography vendor | Index test | Comparator | Reference standard |
|---|---|---|---|---|---|---|
| Balta 202025 | Retrospective cohort study (accuracy of classifying into low and high risk categories) | 17 895 consecutively acquired screening examinations from 1 centre in Germany (0.64% screen detected cancer), age NR | Siemens 70% Hologic 30% | Transpara version 1.6.0 (Screenpoint Medical BV, Nijmegen, Netherlands); preselection of probably normal mammograms; Transpara risk score of 1-10, different cutoff points evaluated; optimal cutoff point ≤7: low risk |
No comparator as human consensus reading decisions used as reference standard for screen negative results | Cancer: biopsy proven screen detected cancers; non-cancer: no information about follow-up for the normal examinations was available. available; for this review, a normal mammogram was defined as free of screen detected cancer based on human consensus reading |
| Dembrower 202026 | Retrospective case-control study (accuracy of classifying into low and high risk categories) | 7364 women with screening examinations obtained during 2 consecutive screening rounds in 1 centre in Sweden (7.4% cancer: 347 screen detected in current round, 200 interval cancers within 30 months of previous screening round), median age 53.6 (IQR 47.6-63.0) | Hologic 100% | Lunit (Seoul, South Korea, version 5.5.0.16) (1) AI for preselection of mammograms probably normal, (2) AI as post-screen after negative double reading to recall women at highest risk of undetected cancer; AI risk score: decimal between 0 and 1, different cutoff points evaluated |
None | Cancer: diagnosed with breast cancer at current screening round or within ≤30 months of previous screening round; non-cancer: >2 years’ follow up |
| Lång 202127 | Retrospective cohort study (accuracy of classifying into low and high risk categories) | 9581 women attending screening at 1 centre in Sweden, consecutive subcohort of Malmö Breast Tomosynthesis Screening Trial (0.71% screen detected cancers), mean age 57.6 (range 40-74) | Siemens 100% | Transpara version 1.4.0 (Screenpoint Medical BV, Nijmegen, Netherlands); preselection of mammograms probably normal; Transpara risk score of 1–10, different cutoff points evaluated; chosen cutoff point ≤5: low risk |
No comparator as human consensus reading decisions used as reference standard for screen negative results | Cancer: histology of surgical specimen or core needle biopsies with a cross reference to a regional cancer register; non-cancer: a normal mammogram was defined as free of screen detected cancer based on human consensus reading |
| Raya-Povedano 202131 | Retrospective cohort study (accuracy of classifying into low and high risk categories) | 15 986 consecutive women from the Córdoba Tomosynthesis Screening Trial, 1 Spanish centre (0.7% cancer: 98 screen detected (FFDM or DBT), 15 interval cancers within 24 months of screening); mean age 58 (SD 6), range 50-69 years | Hologic (Selenia Dimensions) 100% | Transpara, version 1.6.0 (ScreenPoint Medical BV, Nijmegen, Netherlands); preselection of mammograms probably normal; Transpara risk score of 1–10; cutoff point ≤7 low risk (chosen based on previous research by Balta 202035) |
Original radiologist decision from Córdoba Tomosynthesis Screening Trial (double reading without consensus or arbitration) | Cancer: histopathologic results of biopsy, screen detected via FFDM or DBT and interval cancers within 24 months of screening; non-cancer: normal reading with 2-years’ follow-up |
AI=artificial intelligence; DBT=digital breast tomosynthesis; FFDM=full field digital mammography; IQR=interquartile range; NR=not reported; SD=standard deviation.