13 July 2022
Hamm K, Hellingman D, Kotrini L, Vetter B, Jordan T, Entrup C, Engelke M, Janssen N, Schubotz B
AI-based strategy to reduce the recall rate and consensus meeting workload of double reading in breast cancer screening with digital mammography: a retrospective evaluation
To evaluate Transpara when implemented as autonomous artificial intelligence (AI) based triaging strategy in breast cancer screening as compared to independent double reading with consensus.
Materials and Method
A consecutive cohort of 37674 digital mammography screening exams (including 210 screen-detected, 45 interval, and 110 next-round screen-detected (NRSD) cancers) were retrospectively collected from a German screening site. Transpara computed a cancer risk score (from 1 to 10) for each exam. Double reading of all exams was compared with an autonomous AI triaging strategy; no human reading is performed for the least suspicious exams classified by AI (score 1-6), only exams with score 7-10 are double read, and the top 1% most suspicious exams as classified by Transpara are automatically recalled. Cancer detection rate (CDR), recall rate (RR), and consensus workload were evaluated.
Double reading of all exams resulted in a CDR of 5.6/1000, RR of 4.4% (1655/37674), and consensus workload of 4255 exams. A total of 27562 exams (73.2%), including 5 screen detected cancers and 636 false-positive recalls, had a Transpara score of 1-6. Transpara found 5 additional cancers (2 interval cancers and 3 NRSD cancers) and 173 additional false-positive recalls in the 1% most suspicious exams. Autonomous AI triaging based on Transpara would result in a similar CDR (5.6/1000), 28.0% lower RR (down to 3.2%), and 45.1% reduction of consensus workload compared to standard double reading.
Not reading exams with Transpara score 1-6 can reduce radiologists’ RR and workload in screening at the cost of missing some screen-detected cancers. However, recalling the top 1% might compensate this loss in CDR. Transpara-assisted double reading of all exams can potentially lower the RR and increase the CDR, but prospective studies should confirm this.