28 November 2020
C Balta et al
Using AI to triage which screening mammograms benefit from a double reading strategy
C Balta, A Rodriguez-Ruiz, C Mieskes, N Karssemeijer, S H Heywang-Koebrunner
To determine if an artificial intelligence (AI) system can be used to triage which screening digital mammography (DM) exams benefit from a double reading strategy from those exams were single reading could be sufficient.
METHOD AND MATERIALS
A consecutive sample of 17886 screening digital mammograms with 114 biopsy-proven cancers was retrospectively collected. Exams were acquired with Siemens and Hologic DM systems in one screening unit that invites women for biennial screening and uses independent double reading of mammograms. An AI system (Transpara version 1.6.0, ScreenPoint Medical) automatically analyzed each mammogram and assigned an “AIScore” 1-10. The higher the AI-Score the higher the likelihood of malignancy within the DM exam. The AI-Score is calibrated such that in a screening population approximately 10% of mammograms are assigned to each category. The hypothesis was that in DM exams with an AI-Score lower or equal than a threshold, single-reading would have been sufficient, while those exams with higher AI-Score, double-reading would be beneficial. All AI Scores (1-9) were investigated as potential thresholds. Cancer detection rate (CDR), recall rate and positive predictive value (PPV) of the simulated triaging screening strategy were compared to the original outcomes, using a McNemar’s test to analyze paired data.
The original double reading resulted in a CDR of 6.4/1000, recall rate of 5.4% and PPV of 11.9%. By selecting an AI-Score threshold of 7 to triage double reading, the number of exams (n=6229) to be double read could be reduced by 65% without having an impact on sensitivity. CDR would have remained unchanged (6.4/1000), recall rate would have been reduced by 11.8% (down to 4.8%, P<0.001) and PPV would have been increased by 10.5% (up to 13.3%, P<0.001), compared to double reading of all mammograms.
AI can be used to triage screening mammograms that would benefit most from double reading, reducing workload and improving screening performance.
AI-based triaging of screening mammograms for single or double reading can potentially improve the quality of screening and lead to more effective use of radiologists’ reading time on mammograms.