Impact of AI-based Slab Reconstruction Technology on the Diagnostic Accuracy of Screening Digital Breast Tomosynthesis
Menée à partir de l'analyse de 119 662 clichés de tomosynthèses mammaires numériques de dépistage réalisées auprès de 64 949 femmes sur la période 2018-2019 et la période 2021-2022 (âge moyen : 60 ans), cette étude évalue la performance, du point de vue de la sensibilité, de la spécificité, du taux de détection de lésions cancéreuses et du taux de faux négatifs, de la tomosynthèse avant et après l'implémentation d'un modèle d'intelligence artificielle basé sur des technologies de reconstruction par blocs d'images
Background : Digital breast tomosynthesis (DBT) uses 1-mm slices, resulting in a larger number of images and longer interpretation times than conventional digital two-dimensional mammography. Slab reconstruction technologies address this challenge by generating thicker slices, thereby reducing the number of images requiring review, improving efficiency, and lowering storage demand.
Purpose : To compare the diagnostic accuracy of screening DBT before and after the implementation of artificial intelligence (AI)–based slab reconstruction technology.
Materials and Methods : Consecutive screening DBT examinations obtained before and after the implementation of a slab reconstruction technology at an academic medical center were retrospectively reviewed. The slab reconstruction technology uses AI to generate 6-mm synthetic slices with 3-mm overlap and 70-
μm pixel resolution. The preimplementation period was between January 2018 and December 2019, and the postimplementation period was between October 2021 and September 2022. Multivariable logistic regression models were used to compare screening performance metrics in both periods, and a noninferiority analysis was performed.
Results
:
A total of 119
662 screening DBT examinations in 64 949 women were analyzed: 77 577 (52 649 women; mean age, 60 years ± 11 [SD]) during the preimplementation period and 42 085 (42 059 women; mean age, 60 years ± 11) during the postimplementation period. The cancer detection rate (CDR) (5.8 vs 6.5 per 1000 examinations; adjusted odds ratio [OR], 1.1; P = .49), sensitivity (82.3% vs 85.9%; adjusted OR, 1.3; P = .27), and false-negative rate (1.2 vs 1.1 per 1000 examinations; adjusted OR, 0.8; P = .39) did not differ between periods, and all three metrics met the noninferiority criteria. The abnormal interpretation rate (AIR) was lower (6.2% vs 5.8%; adjusted OR, 0.9; P < .001) and the specificity was higher (94.4% vs 94.9%; adjusted OR, 1.1; P < .001) during the postimplementation period.
Conclusion : The implementation of AI-based slab reconstruction technology was associated with noninferior CDR and sensitivity, improved specificity, and reduced AIR.
Radiology , article en libre accès, 2026