European Congress of Radiology 2022 - 16 July 2022

Can an AI-based decision referral approach improve the overall sensitivity of a breast cancer screening program?

- Stefan Bunk et al.

Purpose:

We developed and evaluated an AI system’s decision referral approach, where very confident algorithmic assessments are performed automatically and more difficult assessments are deferred to the radiologist. This two-part system incorporates triage of normal exams, while introducing a “safety net” to maintain high sensitivity by performing predictions on cancer-positive exams.

Methods:

A representative screening program dataset of a total of N=24,501 full-field mammography exams from biopsy-confirmed, screen-detected (n=2,105) plus interval breast cancers (n=2,396), and follow-up proven negatives (n=20,000) from 8 German breast cancer screening units was used to simulate the impact of the decision referral approach on screening program-level sensitivity. Images were analyzed using a commercial AI system. We computed the program-level change in sensitivity based on a previously validated threshold for cancer detection derived from an operating point of 98% screen-detected cancer sensitivity for normal triaging and 99% specificity for the safety net. Upper and lower bounds of the change in sensitivity are computed using resampling methods. Absolute reduction in missed cancers and triaging rate was also calculated.

Results:

The sensitivity of the simulated screening program was 59.7%. At the selected operating point, sensitivity of the screening program is improved by 3.9 (2.8–5.2) percentage points. The AI system detected 20.5% of cancers missed by radiologists, while also offering automation potential with a triaging rate of 50.7% for each radiologist.

Conclusion:

By combining triaging with algorithmic detections of highly suspicious lesions that would otherwise have been missed by radiologists, a decision referral approach demonstrates improvement of sensitivity for the screening program.

Authors:

S. Bunk¹, D. Byng¹, M. Brehmer^1,2, T. Töllner³, L. Umutlu², K. Pinker-Domenig⁴, C. Leibig¹