The potential of AI to replace a first reader in a double reading breast cancer screening program: a feasibility study

Janssen N, Rodriguez-Ruiz A, Mieskes C, Karssemeijer N, Heywang-Köbrunner S H
ECR2021

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

Materials and Method
A randomly selected screening cohort of 2314 DM exams with 114 biopsy-proven screen detected cancers was retrospectively collected from a single center, which involved independent double human reading and consensus in case of disagreement. Normal mammograms were defined as free of screen-detected cancer by double reading. No follow-up information about interval cancers was available. All exams were processed by an AI system (Transpara, ScreenPoint Medical), assigning an Exam Score 1-10 which represents an increasing risk of containing visible abnormalities. An AI-based screening scenario was retrospectively investigated in which the second reader was replaced with Transpara (operating at the same sensitivity). Specificity and workload were compared between original double reading and the AI-based scenario using a McNemar test.

Results
Double human reading recalled 505 women to consensus, with 112 detected cancers and 393 false positives (specificity 82.1%, CI: 80.5-83.7). AI-based screening resulted in 550 recalled women to consensus, with 112 detected cancers and 438 false positives (specificity 80.1%, CI: 78.4-81.7, relative reduction compared to double reading of 2.0%, CI: -0.3-4.3%, P=0.08).

Conclusion
Replacing a human reader with Transpara in a double reading breast cancer screening program could be an effective strategy to halve radiologist’s workload in screening while maintaining the same sensitivity, particularly in case of a shortage of readers. This strategy potentially comes at the cost of more false positive cases sent to consensus.