Influence of AI Decision Support on Radiologists' Performance and Visual Search in Screening Mammography

Background: Artificial intelligence (AI) decision support may improve radiologist performance during screening mammography interpretation, but its effect on radiologists’ visual search behavior remains unclear. Purpose: To compare radiologist performance and visual search patterns when reading screening mammograms with and without an AI decision support system. Materials and Methods: In this retrospective multireader multicase study, 12 breast screening radiologists with 4–32 years of experience (median, 12 years) from 10 institutions evaluated screening mammograms acquired between September 2016 and May 2019. Assessments were conducted unaided and with a Food and Drug Administration–approved, European Commission–marked AI decision support system, which assigns a region suspicion score from 1 to 100, with 100 indicating the highest malignancy likelihood. An eye tracker monitored readers’ eye movements. Area under the receiver operating characteristic curve (AUC), sensitivity, and specificity between unaided and AI-assisted reading were compared using multireader multicase analysis software. Reading times, breast fixation coverage (percentage breast covered by fixations within 2.5° visual angle radius) fixation time, and time to first fixation within the lesion region were compared using bootstrap resampling (n = 20 000). Results: Mammography examinations (75 with breast cancer, 75 without breast cancer) from 150 women (median age, 55 years [IQR, 50–63 years]; age range, 49–72 years) were read. The mean AUC was higher with AI support versus unaided reading (unaided, 0.93 [95% CI: 0.91, 0.96]; AI-supported, 0.97 [95% CI: 0.95, 0.98]; P < .001). There was no evidence of a difference in mean sensitivity (81.7% [735 of 900 readings] vs 87.2% [785 of 900]; P = .06), specificity (89.0% [801 of 900] vs 91.1% [820 of 900]; P = .46), or reading time (29.4 vs 30.8 seconds; P = .33). Breast fixation coverage was lower with AI support (11.1% vs 9.5% of breast area; P = .004), while fixation time in the lesion region was higher (4.4 vs 5.4 seconds; P = .006). There was no evidence of a difference in time to first fixation within the lesion region (3.4 vs 3.8 seconds; P = .13). Conclusion: Radiologists improved their breast cancer detection accuracy when reading mammography with AI support, spending more fixation time on suspicious areas and less on the rest of the breast, indicating a more efficient search.