Can artificial intelligence identify normal mammograms in screening?

Lång K, Dustler M, Dahlblom V, Andersson I, Zackrisson S

Aim and Objective
This study was performed to see whether it was possible to identify normal mammograms in screening using Artificial Intelligence (AI). It determined whether AI can reduce a Radiologist’s workload of reading normal exams and assessed the effect on false positives.

Materials and Method
This was an ethically approved retrospective study using a screening population. Consecutive inclusion of mammograms resulted in the use of 9,581 double-read mammography screening exams.
In this cohort, there were 68 screen-detected cancers and 187 false positives (recall rate of 2.7% and a cancer detection rate of 7.1 per 1,000 exams).
The deep learning-based AI CAD system used was Transpara™ AI v1.4.0, ScreenPoint Medical, Nijmegen, Netherlands which has a pre-selection tool categorizing mammograms into a risk score of between 1 and 10.

Results
Transpara™ AI analysis of the 9,581 2D exams demonstrated that:
– 5,082 (53%) had low-risk scores. This included 7 cancers and 52 false positives.
– By excluding risk scores 1 and 2, it is possible to remove 1,829 screen exams from the workload without missing a single cancer – a 19% reduction in screen reading.
– 10 false positives could have been avoided – a 5% reduction.

Conclusion
The study results demonstrated that using Transpara™ AI in mammography screening could result in 19% of screening mammograms being excluded from Radiologist’s worklists, without missing a single cancer. In addition, 5% of false positives could be avoided by excluding normal screen exams.
These findings indicate a strong potential for cost reduction related to screen reading and the reduction in false positives.
By adapting the screening workflow into 3 risk groups; AI only (0% cancers/ 19% of the screening population), single read only (31% of cancers/ 69% of the screening population), and double read (69% of cancers/ 12% of the screening population), it is possible to reduce the screening workload by 54%.