AI-decision support for double reading in breast cancer screening with digital mammography: A pseudo-prospective evaluation

Purpose

To evaluate the added value of artificial intelligence (AI) as decision support in a region of the German breast cancer screening program with independent double reading and consensus.

Methods and Materials

Between January and September 2022, a cohort of 21.030 digital mammography screening exams (including 118 screen-detected cancers) was double read with AI support by 3 readers. Transpara (version 1.7.2) provides a risk category (low, intermediate or elevated) for each exam, indicating the likelihood of malignancy. Readers consulted Transpara whenever they were uncertain whether to forward an exam to consensus. Moreover, Transpara was used as decision support during consensus.

The screening performance was compared to the performance of a historic cohort before the implementation of Transpara, including 22 656 screening exams (of those, 94 cancers), recorded between January and September 2021, and read by the same 3 readers. The Transpara risk categories were retrospectively retrieved. The amount of cancers per risk category was calculated. Screening performance was evaluated in terms of cancer detection rate (CDR) , recall rate (RR), consensus workload, and positive predictive value (PPV).

Results

Overall, 99.1% (117/118) of screen-detected cancers were classified by Transpara as being intermediate or elevated risk. Implementation of Transpara as decision support led to 26.8% increase in CDR (no AI: 4.1/1000; with AI: 5.6/1000, p = .028), going along with a slight increase in RR (no AI: 2.2%, with AI:  2.5%, p = .04) and PPV (no AI: 18.7%, with AI:  22.3%, p = .04). 

Conclusion

The implementation of AI-based decision support in a double reading setting significantly improved the quality of breast cancer screening in terms of cancer detection.