Using AI with single reading in screening: a simulation of the impact on tumour characteristics of detected cancers

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

Aim and Objective
To compare the tumor characteristics of screen-detected cancers when one human reader is replaced by an artificial intelligence (AI) system in a double reading breast cancer screening program.

Materials and Method
A consecutive cohort of 23,035 digital mammography exams with 159 screen-detected (SD), 48 interval (IC) and 62 next-round-SD cancers (NRSD) was collected from the Dutch screening program. An AI system (Transpara, ScreenPoint Medical) processed all exams and assigned a recall/no recall decision at the same recall rate as a radiologist (3.0%).
Histopathological types (invasive ductal/lobular carcinoma (IDC/ILC), ductal carcinoma in situ (DCIS)), grade, and mean radiological size (95% CI) of all cancers were reported for double reading and single reading + AI.

Results
Double reading detected 159 SD, with mean size of 13.6mm (95% CI:12.4-14.9mm). These were 61.0% IDC, 10.1% ILC, 19.5% DCIS, 7.5% other invasive and 1.9% unknown. Single reading + AI recalled 179 cancers (154 SD, 15 IC, 10 NRSD), with mean size 14.2mm (95% CI:13.1-15.4mm) and 62% IDC, 12.8% ILC, 16.2% DCIS, 7.3% other invasive and 1.7% unknown. Double reading and single reading + AI detected similar proportions of grade-3 cancers: 25% vs. 21.8% (p=0.48). The five SD that were missed by single reading + AI were all grade-3, (3 DCIS, 1 ILC, 1 IDC). The 15 recalled IC by single reader + AI were all invasive, and 20% were grade-3, 33% grade-2, 40% grade-1 and 7% unknown. Of the 10 NRSD, 10% were grade-3, 20% grade-2, 60% grade-1 and 10% unknown, with 1 DCIS.

Conclusion
Introducing Transpara in screening could lead to higher cancer detection rates with similar tumour characteristics as with double reading of mammograms.