Artificial Intelligence in Mammography Screening in Norway

Independent double reading with consensus is the recommended practice for breast cancer screening programs in Europe. This is a time-consuming process and more than 99% of examinations are determined to have a negative final outcome in BreastScreen Norway. However, radiologist still overlook cancers, and review-studies have shown 20-25% of screen-detected and interval cancers to be visible at prior screening mammograms. In Norway and other countries, there is also a shortage of breast radiologists.

Retrospective studies on AI in mammographic screening have shown promising results in the classification of cancer negative and cancer positive examinations. However, prospective studies are needed to elucidate questions concerning AI as a support or replacement for the radiologists in the interpretation process.

In this randomized controlled trial women attending screening and consenting to participate in the study, will be randomized (1:1 allocation) to an intervention or a control group. The screening examination will be performed as usual in both groups. The intervention in the trial is the interpretation procedure.

All examinations will be analyzed by AI. In the intervention group, examinations will be triaged according to risk of malignancy defined by an AI score. Examinations with a low AI score (1-7) will have a single reading by one radiologist and examinations with a intermediate or high AI score (8-10) will have an independent double reading by two radiologists. All examinations selected by either one or both radiologist will be discussed at consensus. Examinations with an AI raw score of \>9.8 will be discussed at consensus independent of the radiologists' interpretation. The radiologists will be blinded to whether the mammograms are interpreted by one or two radiologists. AI risk scores and AI image marking of suspicious calcifications and masses are only provided to the radiologists at consensus, not during independent reading. In the control group, two independent radiologists will interpret all screening examinations according to standard procedure, independent double reading. AI results will not be available at either independent reading or consensus, only for comparison with the intervention group in final analyses.

Screen-detected cancers in the intervention and control arm will be analyzed in terms of non-inferiority.

• Consensus rate [Срок оценки: 24 months from study start-up]
• Recall rate [Срок оценки: 24 months from study start-up]
• Interval breast cancers [Срок оценки: 60 months from study start-up]
• Prognostic and predictive tumor characteristics of screen-detected breast cancer [Срок оценки: 36 months from study start-up]
• Prognostic and predictive tumor characteristics of interval breast cancer [Срок оценки: 60 months from study start-up]
• Sensitivity [Срок оценки: 60 months from study start-up]
• Specificity [Срок оценки: 60 months from study start-up]
• Mammographic features of screen-detected breast cancer [Срок оценки: 36 months from study start-up]
• Time spent screen-reading [Срок оценки: 36 months from study start-up]