Reading breast tomosynthesis examinations with an AI decision support system: improving cancer detection accuracy

Mann R, Rodriguez-Ruiz A, Gubern-Merida A, Karssemeijer N, Sechopoulos I.

Purpose:

To compare the breast cancer detection accuracy of radiologists reading breast tomosynthesis (DBT) examinations unaided versus supported by an artificial intelligence (AI) system.

Methods:

A cancer-enriched, retrospective, fully-crossed, multi-reader, multicase, HIPAA-compliant study was performed. Four-view DBT examinations from 240 women were included. All examinations (in total 71 breasts with cancer lesions, 70 breasts with benign findings, and 339 normal breasts) were interpreted by 9 qualified radiologists (median experience 8 years, range 4-23), once with and once without AI support (Transpara, ScreenPoint Medical). The readers provided a level of suspicion for each breast. When using AI support, radiologists were shown an examination-based cancer likelihood score as well as marked lesions and lesion-based cancer likelihood scores. The area under the receiver operating characteristic curve (AUC) was compared between both reading conditions at a per breast-level, using mixed-models analysis of variance for multiple repeated clustered measurements. Reading time differences for normal examinations were also measured.

Results:

On average, the AUC was higher with AI support than with unaided reading (0.861 vs 0.820, respectively; P = .001). The AUC of the stand-alone AI system was similar to the average AUC of the radiologists unaided (0.840 vs 0.820, 95% CI: -0.038, +0.078). Reading time per normal case showed an average reduction of -21% when using AI, with 8 of 9 readers having shorter reading time.

Conclusion:

Radiologists improved their cancer detection in breast tomosynthesis examinations when using an AI system for support, while simultaneously reducing reading time.