Accuracy
Risk

ECR 2023

1 March 2023

Authors

HW Koch, M Larsen, H Bartsch, S Hofvind

Artificial intelligence (AI) and extremely dense breasts in BreastScreen Norway: Can AI increase cancer detection?

Purpose

Women with mammographic extremely dense breasts have a 3-6 times higher risk of developing breast cancer compared to women with fatty breasts. Extremely dense breasts represent a challenge for radiologists due to the masking effect of tumours. In March 2022, the European Society of Breast Imaging (EUSOBI) recommended to offer women with extremely dense breasts an MRI-based screening every 2-4 years. We consider this recommendation to be very costly and would like to propose other options for dealing with this problem. The aim of this study was to compare the performance of an AI system in extremely dense breast tissue versus independent double reading by radiologists.

Methods and Materials

In this retrospective study, the study sample included 67 screen-detected cancers, 38 interval cancers and 2403 negative screening examinations among women with extremely dense breasts. The 2508 women were screened in BreastScreen Norway during the period 2010 to 2018. The AI-system Transpara scored examinations on a scale from 1 to 10, based on the risk of malignancy. Sensitivity for the AI system was calculated by setting AI score 10 as the threshold for a true positive screening examination. When calculating sensitivity for independent double reading, we considered all interval cancers as false negative.

Results

All 67 screen-detected cancers and 50.0% of the interval cancers had an AI score of 10. A total of 16.1% of the negative examinations had an AI score of 10. Sensitivity for independent double reading was 63.8% and 81.9% for Transpara.

Conclusion

Our results indicate promising performance of Transpara for women with extremely dense breasts. The cost of an increased number of false positive cases selected for consensus needs to be further explored in prospective studies.

Limitations

Due to the retrospective nature of the study, this dataset does not reflect the actual screening setting and the actual number of cases with a high AI score that would be detected in a prospective screening setting cannot be determined.


You might also like