Retrospective comparison between single reading plus an artificial intelligence algorithm and two-view digital tomosynthesis with double reading in breast screening.

Abstract

To examine the breast cancer detection rate by single reading of an experienced radiologist supported by an artificial intelligence (AI) system, and compare with two-dimensional full-field digital mammography (2D-FFDM) double reading. Images (3D-tomosynthesis) of 161 biopsy-proven cancers were re-read by the AI algorithm and compared to the results of first human reader, second human reader and consensus following double reading in screening. Detection was assessed in subgroups by tumour type, breast density and grade, and at two operating points, referred to as a lower and a higher sensitivity threshold. The AI algorithm method gave similar results to double-reading 2D-FFDM, and the detection rate was significantly higher compared to single-reading 2D-FFDM. At the lower sensitivity threshold, the algorithm was significantly more sensitive than reader A (97.5% vs. 89.4%, p = 0.02), non-significantly more sensitive than reader B (97.5% vs. 94.4%, p = 0.2) and non-significantly less sensitive than the consensus from double reading (97.5% vs. 99.4%, p = 0.2). At the higher sensitivity threshold, the algorithm was significantly more sensitive than reader A (99.4% vs. 89.4%, p < 0.001) and reader B (99.4% vs. 94.4%, p = 0.02) and identical to the consensus sensitivity (99.7% in both cases, p = 1.0). There were no significant differences in the detection capability of the AI system by tumour type, grading and density. In this proof of principle study, we show that sensitivity using single reading with a suitable AI algorithm is non-inferior to that of standard of care using 2D mammography with double reading, when tomosynthesis is the primary screening examination.