The potential for AI to replace a reader in a double reading breast cancer screening program

van Winkel S, Janssen N, Rodriguez-Ruiz A, Karssemeijer N, Sechopoulos I, Mann R M
ECR2021

Aim and Objective
To evaluate the impact of replacing the double human reading with a single human reading and an artificial intelligence (AI) system in a breast cancer screening program.

Materials and Method
A consecutive cohort of 23,035 digital mammography (DM) exams with 159 screen detected (SD), 48 interval (IC) and 62 next-round SD cancers (NRSD) was collected from the Dutch screening program, which involved double human reading. An AI system (Transpara, ScreenPoint Medical) processed all cases, assigning a level of suspicion to each exam on a continuous scale. AI-based screening was retrospectively investigated in which the second reader was replaced with Transpara, using the human recall rate as operating point. Sensitivity and recall rate were computed for the first reader, Transpara alone, AI-based screening and double reading.

Results
The first reader recalled 694 cases (3.0%, CI = 2.8-3.2%), including 138 SD, 2 IC and 1 NRSD (sensitivity 52.4%, CI =46.3-58.5%). AI alone detected 127 SD, 14 IC and 10 NRSD (sensitivity 56.1%, CI = 50.0-62.2%). There were 1012/23035 (4.4%) discrepant assessments between AI and the first reader. AI-based screening recalled 1200 cases (5.2%, 95% CI = 4.9-5.5%), detecting 154 SD, 15 IC and 10 NRSD (sensitivity 66.5%, CI =60.6-72.2%). Double reading (including consensus) recalled 724 cases (3.1%, CI = 2.9-3.4%), with 159 SD (sensitivity 59.1%, CI = 53.0-65.0%).

Conclusion
The use of Transpara as an independent reader could halve the radiologist’s workload and potentially increase sensitivity and lower recall rate, if an effective arbitration process can be implemented to assess discrepant exams.