Population-wide evaluation of artificial intelligence and radiologist assessment of screening mammograms.

OBJECTIVES: To validate an AI system for standalone breast cancer detection on an entire screening population in comparison to first-reading breast radiologists. MATERIALS AND METHODS: All mammography screenings performed between August 4, 2014, and August 15, 2018, in the Region of Southern Denmark with follow-up within 24 months were eligible. Screenings were assessed as normal or abnormal by breast radiologists through double reading with arbitration. For an AI decision of normal or abnormal, two AI-score cut-off points were applied by matching at mean sensitivity (AIsens) and specificity (AIspec) of first readers. Accuracy measures were sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and recall rate (RR). RESULTS: The sample included 249,402 screenings (149,495 women) and 2033 breast cancers (72.6% screen-detected cancers, 27.4% interval cancers). AIsens had lower specificity (97.5% vs 97.7%; p < 0.0001) and PPV (17.5% vs 18.7%; p = 0.01) and a higher RR (3.0% vs 2.8%; p < 0.0001) than first readers. AIspec was comparable to first readers in terms of all accuracy measures. Both AIsens and AIspec detected significantly fewer screen-detected cancers (1166 (AIsens), 1156 (AIspec) vs 1252; p < 0.0001) but found more interval cancers compared to first readers (126 (AIsens), 117 (AIspec) vs 39; p < 0.0001) with varying types of cancers detected across multiple subgroups. CONCLUSION: Standalone AI can detect breast cancer at an accuracy level equivalent to the standard of first readers when the AI threshold point was matched at first reader specificity. However, AI and first readers detected a different composition of cancers. CLINICAL RELEVANCE STATEMENT: Replacing first readers with AI with an appropriate cut-off score could be feasible. AI-detected cancers not detected by radiologists suggest a potential increase in the number of cancers detected if AI is implemented to support double reading within screening, although the clinicopathological characteristics of detected cancers would not change significantly. KEY POINTS: • Standalone AI cancer detection was compared to first readers in a double-read mammography screening population. • Standalone AI matched at first reader specificity showed no statistically significant difference in overall accuracy but detected different cancers. • With an appropriate threshold, AI-integrated screening can increase the number of detected cancers with similar clinicopathological characteristics.

Breast cancer detection accuracy of AI in an entire screening population: a retrospective, multicentre study.

BACKGROUND: Artificial intelligence (AI) systems are proposed as a replacement of the first reader in double reading within mammography screening. We aimed to assess cancer detection accuracy of an AI system in a Danish screening population. METHODS: We retrieved a consecutive screening cohort from the Region of Southern Denmark including all participating women between Aug 4, 2014, and August 15, 2018. Screening mammograms were processed by a commercial AI system and detection accuracy was evaluated in two scenarios, Standalone AI and AI-integrated screening replacing first reader, with first reader and double reading with arbitration (combined reading) as comparators, respectively. Two AI-score cut-off points were applied by matching at mean first reader sensitivity (AIsens) and specificity (AIspec). Reference standard was histopathology-proven breast cancer or cancer-free follow-up within 24 months. Coprimary endpoints were sensitivity and specificity, and secondary endpoints were positive predictive value (PPV), negative predictive value (NPV), recall rate, and arbitration rate. Accuracy estimates were calculated using McNemar's test or exact binomial test. RESULTS: Out of 272,008 screening mammograms from 158,732 women, 257,671 (94.7%) with adequate image data were included in the final analyses. Sensitivity and specificity were 63.7% (95% CI 61.6%-65.8%) and 97.8% (97.7-97.8%) for first reader, and 73.9% (72.0-75.8%) and 97.9% (97.9-98.0%) for combined reading, respectively. Standalone AIsens showed a lower specificity (-1.3%) and PPV (-6.1%), and a higher recall rate (+ 1.3%) compared to first reader (p < 0.0001 for all), while Standalone AIspec had a lower sensitivity (-5.1%; p < 0.0001), PPV (-1.3%; p = 0.01) and NPV (-0.04%; p = 0.0002). Compared to combined reading, Integrated AIsens achieved higher sensitivity (+ 2.3%; p = 0.0004), but lower specificity (-0.6%) and PPV (-3.9%) as well as higher recall rate (+ 0.6%) and arbitration rate (+ 2.2%; p < 0.0001 for all). Integrated AIspec showed no significant difference in any outcome measures apart from a slightly higher arbitration rate (p < 0.0001). Subgroup analyses showed higher detection of interval cancers by Standalone AI and Integrated AI at both thresholds (p < 0.0001 for all) with a varying composition of detected cancers across multiple subgroups of tumour characteristics. CONCLUSIONS: Replacing first reader in double reading with an AI could be feasible but choosing an appropriate AI threshold is crucial to maintaining cancer detection accuracy and workload.