Radiology 2022
29 März 2022
Authors
Larsen M, Flåt Aglen C, Lee C, Roth Hoff S, Lund-Hanssen H, Lång K, Nygård J, Ursin G, Hofvind S
Artificial Intelligence Evaluation of 122 969 Mammography Examinations from a Population-based Screening Program
Background:
Artificial intelligence (AI) has shown promising results for cancer detection in mammographic screening. However, evidence related to the use of AI in real screening settings remains sparse.
Objective:
We compared Transpara’s performance to routine independent double reading with consensus as usually performed in a population-based screening program. Further, we explored histopathological characteristics of tumors with different Transpara scores.
Materials and Methods:
In this retrospective study, 122 969 screening examinations from 47 877 women performed at four facilities of BreastScreen Norway from October 2009 to December 2018 were included. The dataset included 752 (6.1/1000) screen-detected and 205 (1.7/1000) interval cancers. Each examination had an Transpara examscore between 1 and 10, with 1 indicating a low risk of breast cancer and 10 indicating a high risk. The thresholds T1, T2, and T3 were used to assess Transpara’s performance as a binary decision tool (selected versus not selected). T1 was set at Transpara score 10, T2 was chosen to yield a selection rate similar to the consensus rate (8.8%) and T3 to yield a selection rate similar to an average individual radiologist (5.8%).
Results:
A total of 653 out of 752 (86.8%) screen-detected and 92 out of 205 (44.9%) interval cancers received a Transpara score of 10 (T1). Using T3, 80.1% (602 of 752) of the screen-detected and 30.7% (63 of 205) of the interval cancers were detected. Screen-detected cancer with Transpara scores not selected using the three thresholds had favorable histopathological characteristic compared to those selected; opposite results were observed for interval cancers.
Conclusion:
The proportion of screen-detected cancers not selected by Transpara at three evaluated thresholds was less than 20%. According to cancer detection, the overall performance of Transpara was promising.