Performance benchmarking


29 November 2020


Marta Pinto et al

Impact of Artificial Intelligence decision support on breast cancer screening interpretation with single-view wide-angle digital breast tomosynthesis

Marta Pinto, Alejandro Rodriguez-Ruiz, Kristin Pedersen, Solveig Hofvind, Ritse Mann, Steffen Kappler, Julia Wicklein, Ioannis Sechopoulos.

Presented RSNA2020

To compare radiologist’s accuracy when reading single-view wide-angle digital breast tomosynthesis (DBT) images with and without an artificial intelligence (AI) system for decision and navigation support.

Materials and Methods
A multi-reader, multi-case observer study was performed on bilateral medio-lateral oblique (MLO) cases acquired with a wide-angle DBT system and their corresponding synthetic 2D images (software v. VC20A, Mammomat Inspiration, Siemens). Eight breast radiologists (median screening experience: 5.25 years, range: 3-15 years) interpreted 190 cases (90 normal, 26 benign, 74 malignant), including both DBT stack and synthetic, with ground truth verified by histopathological analysis or two-year follow-up. Reading was performed in two sessions, separated by at least 4 weeks, with a random mix of cases read with and without AI decision and navigation support (Transpara, ScreenPoint) in each session by each reader. When reading with AI support, readers were directed to the lesion-containing slice when clicking on the AI finding marked on the synthetic. Forced BI-RADS® (1-5) and level of suspicion (1-100) scores were given per breast by each reader. The area under the receiver operating characteristics curve (AUC) and the specificity and sensitivity were compared between conditions using the iMRMC software. The average reading time per reader was compared using the Wilcoxon signed-rank test.

The per case AUC for the average reader was higher when interpreting images with AI support than when reading unaided (0.878 vs 0.843, respectively; p = 0.0048). Sensitivity increased with AI support (0.875 vs 0.821; p = 0.0126), while no statistically significant differences in specificity (0.676 vs 0.694; p = 0.3938) or reading time (50 s vs 45 s; p = 0.2597) were detected. However, with AI support, readers spent less time on cases assessed low-suspicion scores by the AI system and more time in high-suspicion cases.

Radiologists improved their cancer detection performance at DBT when using an AI system for decision support. Furthermore, AI support results in a more optimal distribution of reading time, with readers spending less time in normal cases and more time in suspicious cases.

Clinical Relevance/Application
Improvement in breast cancer screening due to the addition of an AI system for decision and navigation support may improve screening outcomes, specifically by reducing the interval cancer rate.

You might also like