Evaluation of the performance of artificial intelligence (AI) after one year of use in breast cancer screening practice: is the promise being delivered?

Purpose
To evaluate the impact of using Transpara as support for human double reading in a breast cancer screening program and its ability to correctly stratify these exams according to probability of cancer.

Methods and materials
We reviewed all digital mammography (DM) or digital breast tomosynthesis (DBT) screening examinations between March 2021 and March 2022 that were double read (without consensus) by radiologists concurrently with AI support at our hospital. Transpara categorises each exam into three categories (low, intermediate, elevated) representing the probability of cancer and highlights suspicious areas (1-100 score). We computed the number of examinations, cancers, recalls and positive predictive value (PPV) of the recalled studies globally and in each AI category, as well as the overall cancer detection rate (CDR) and recall rate (RR) during the study period. We compared these data with the same 12 months period one year earlier (CDR 5.5/1000, RR 6.1%, PPV 9%), prior to the implementation of AI, using the Chi2 test.

Results
11998 screening examinations were included and classified as low: 7917 (65.9%), intermediate: 3730 (31%) and elevated: 351 (2.9%). 108 cancers were detected, which were categorised as low: 1 (0.9%), intermediate: 32 (29.6%), elevated: 75 (69.4%). AI correctly marked 101 cancers. CDR, RR and PPV were 9/1000 (+3.5/1000, p<0.001); 6.1% (0%, p=0.9) and 14.6% (+5.6%, p<0.001), respectively.

Conclusion
Using Transpara concurrently in clinical practice allows to stratify examinations according to probability of cancer. Transpara increases cancer detection rate and positive predictive value of the recalled women.