Can we reduce the workload of mammographic screening by automatic identification of normal exams with artificial intelligence? A feasibility study.

Rodriguez-Ruiz A, Lång K, Gubern-Merida A, Teuwen J, Broeders M, Gennaro G, Clauser P, Helbich T.H, Chevalier M, Mertelmeier T, Wallis M.G, Andersson I, Zackrisson S, Sechopoulos I, Mann R.M
European Radiology (2019). https://doi.org/10.1007/s00330-019-06186-9

Aim and Objective
To study the feasibility of automatically identifying normal digital mammography (DM) exams with artificial intelligence (AI) to reduce the breast cancer screening reading workload.

Materials and Method
Multi-vendor (Siemens, GE, Hologic, Philips) and multi-center (collected at 9 different sites across Europe and USA) sample of DM exams: total of 2652 DM exams. 653 biopsy proven cancer, 768 false positive / benign cases and 1231 normal. Each exam was read by at least 3 different radiologists (28,296 interpretations by 101 radiologists) – who assign a level of suspicion or BI-RADS score and analyzed by Transpara (version 1.4.0), yielding the Transpara Score (TS 1-10) for each exam. The following simulation was performed:

Cases with Transpara Score (TS) higher than X were read by humans, while if they were lower or equal than X they would be automatically assigned as “normal” and not
recalled.

Where X = all possible thresholds of TS 1 to 9

For each threshold X, the estimated impact was measured in terms of:
Screening workload, false positives (specificity) and true positives (sensitivity)

Results
Two thresholds were the most interesting:
If cases with TS 5 or lower would not be read by humans:
– Workload would be approximately halved (− 47%)
– False positives are reduced by 28% and true positives decrease by 7% (same accuracy)
If cases with TS 2 or lower would not be read by humans:
– Workload would be approximately reduced by 20%
– False positive are reduced by 10% and true positives decrease by 1% (same accuracy)

Conclusion
It is possible to automatically pre-select exams using Transpara to significantly reduce the breast cancer screening reading workload. Excluding exams with the lowest likelihood of cancer (TS2) in screening may not change radiologists’ breast cancer detection performance. By excluding exams with the lowest likelihood of cancer (TS1-5), the decrease in true-positive recalls would be balanced by a simultaneous reduction in false-positive recalls.