Reduced workload for breast radiologists: results from a retrospective study using artificial intelligence in mammographic screening

Purpose
Transpara has the potential to reduce radiologists’ workload in mammographic screening without reducing cancer detection. Studies with retrospective data have reported
promising results, but there are limitations concerning small numbers of total exams and enriched datasets. We aimed to identify a scenario where cancer detection remained the same as for standard independent double reading, but where the radiologists’ workload for initial interpretation was reduced.

Materials and Method
Retrospective data from 122,969 screening examinations performed at two breast centres in BreastScreen Norway were used, acquired between 2009 and 2018. The mammograms were processed with the AI-system Transpara (version. 1.7.0), developed by ScreenPoint Medical. Transpara classified screening examinations according to a 10-point scale (AI-score), with a score of 10 reflecting the highest probability of breast cancer.

Results
A total of 78% (745/952) of the cancers had an AI-score of 10, including 87% (653/752) of screen-detected and 45% (92/205) of interval cancers. In a scenario where examinations with an AI-score 1-5 were read by one radiologist (random selection among the two radiologists) and those with a score 6-10 were read by two independent radiologists, the cancer detection rate was similar to the rate observed in a real screening setting: 6.1/1000 screening examinations. This scenario would reduce the radiologists’ workload by 25%.

Conclusion
We found in a retrospective analysis that Transpara as a replacement for a second radiologist in a double reading setting would have decreased the radiologist workload without impacting cancer detection. However, prospective studies are needed to identify the optimal combinations of Transpara and radiologists in double reading setting.