Artificial Intelligence as a support to the radiologists’ screen reading of mammograms – A retrospective study

AIM AND OBJECTIVE
To explore the performance of an artificial intelligence (AI)-system on cancer detection in a population based screening program using independent double reading with consensus, and describe histopathological characteristics of the cancers detected.

MATERIALS AND METHOD
We included data from 65,983 screening examinations including 421 screen-detected and 114 interval cancers, resulting in a program sensitivity of 79% (421/535). The screening examinations were performed as a part of the regular screening setting in two counties in BreastScreen Norway, between 2014 and 2018. We retrospectively processed the mammograms using an AI-system (Transpara, ScreenPoint Medical) which gives each examination a score from 1 to 10. We present the number of breast cancers that were given a threshold value of 9.01 (score=10) and 9.63 by the AI-system. The latter corresponded to a recall rate of 3.5%, which represent the average for the population studied. Median tumor diameter and the proportion of lymph node positive tumors (LN+) were described for cancers with a score of 10 and a score below 10.

RESULTS
Transpara scored 10 for 81% (341/421) of the screen-detected and 41% (47/114) of the interval cancers. Using a threshold of 9.63, 70% (294/421) of the screen-detected and 20% (23/114) of the interval cancers were marked. Screen-detected cancers with a Transpara score of 10 had a median tumor diameter of 13mm (IQR: 9-20) and 18% were LN+, while those with a score below 10 had a median diameter of 10mm (IQR: 7-18) and 12% were LN+. Interval cancers with a Transpara score of 10 had a median tumor diameter of 18m (IQR: 13-25) while those with a score below 10 were 15mm (IQR: 10-20).

CONCLUSION
Transpara marks a substantial number of the screen-detected and interval cancer and could potentially aid radiologists in their screen-reading and increase the sensitivity of the screening program.