The potential of AI for improving early detection in breast cancer screening to reduce interval cancer rates

Alexander J.T. Wanders, Alejandro Rodriguez-Ruiz, Willem Mees, Ioannis Sechopoulos, Nico Karssemeijer, Jan-Kees van Rooden, Ritse M. Mann

Purpose:

To determine the accuracy of an artificial intelligence (AI) system to detect signs of cancer in screening mammograms of interval cancer (IC) cases.

Materials and Methods:

A retrospective case-control study was performed. The last screening mammograms of all IC cases diagnosed in the region South West of the Dutch Breast Screening Program between 2011 and 2015 were consecutively collected. During this period, recall, cancer detection, and interval cancer rates varied per year between 2.1-2.5%, 5.9-6.9/1000 and 2.2-2.6/1000, respectively. The histopathology, pTNM Pathological Classification, and treatment information were retrieved when available. For each IC case, two control cases (normal mammograms verified by at least 2-year follow-up) were collected from the same screening program, matched by age and year of examination. All mammograms were processed by an AI cancer detection system (Transpara, ScreenPoint Medical). Breast density was computed using automated software (LIBRA, Univ. of Pennsylvania). The performance of the AI system for detection of cancer in IC cases at screening prior to diagnosis was estimated in terms of the area under the receiver operating characteristic curve (AUC).

Results:

In total 2332 IC cases (median 15 months from negative screening until IC diagnosis, interquartile range 9-20 months) and 4664 controls were included. The AUC of AI to detect IC in screening was 0.73 (95% CI = 0.72-0.74). The sensitivity was 37.1% and 18.5% at specificity of 90% and 97.5%, respectively. In comparison, the AUC of breast density was 0.70 (95% CI = 0.68-0.73), with sensitivity of 23.3% and 6.5% at specificity of 90% and 97.5%. No difference in AI AUC was found when stratifying by DCIS/invasive IC. A small trend towards higher AI AUC was found for IC diagnosed after an incident round (0.74 vs 0.70 after a prevalent round), as well as for IC treated with breast amputation (0.75 vs 0.72 for IC that received breast-conserving treatments).

Conclusion:

AI has the potential to reduce interval cancer rates, whether by being used as an independent reader or as a pre-selection tool to determine which cases could benefit from additional screening imaging.

Clinical relevance:

There is still a stable rate of interval cancers (often with worse prognosis than screening-detected cancers) in screening programs. New AI-based strategies could help to reduce interval cancer rates.