Diagnostic Performance of AI for Cancers Registered in A Mammography Screening Program: A Retrospective Analysis


Köylüoğluu Y. O. , Şeker M. E. , Arıbal M. E.

ECR 2021, European Congress of Radiology, Vienna, Austria, 3 - 07 March 2021

  • Publication Type: Conference Paper / Unpublished
  • City: Vienna
  • Country: Austria

Abstract

8758 women participated in a 10-year screening program and a subset of participants were included in this study: From a total 110 cancer cases, 74 of them were found during screening, 27 of them were interval cancers, that were detected at a minimum of 4/22/90 days, a maximum of 726 days and a median of 234.5 days and finally 9 of them were missed cancers. There were also 101 healthy mammograms taken from patients who were free from cancer for at least 2 years and was used as a control group. 

We compared and analyzed three different mammography assessment methods in this study: 1) two radiologists assessment at screening center, 2) AI assessment based on the established risk score threshold 3) a hypothetical radiologist and AI team-up where AI is the secondary reader to radiologists. 

Threshold for risk scores, above which were considered to be cancerous was established at 34.5% with a sensitivity and specificity of 72.8% and 88.3% respectively, with an area under curve of 0.853 (95% CI = 0.801 – 0.905). This threshold was established with Youden’s Index, which denotes the point that provides the greatest sensitivity + specificity total. 

According to this threshold, 83.8% of previously detected cancers, 44.4% of interval cancers and 66.6% missed cancers would have been detected by the AI. 

  • Radiologists detected 67.3% of all cancers and missed 32.7%. Missed cancers included some otherwise very easy to detect cancers which would have been detected under normal circumstances.

  • AI detected 72.7% of all cancers, out of which 56.4% were previously detected, 10.9% were interval cancers and 5.5% were missed cancers.

  • Radiologist and AI team-up detected 83.6% of all cancers compared to 67.3% and 72.7% of radiologists and AI on their own, respectively. This approach combines AI’s ability to detect interval and missed cancers with radiologists’ assessment capabilities and results in a greater amount of cases that are detected earlier during screening.