Hospital of the University of Pennsylvania

Traditional 2D computer-aided detection (CAD) showed limited or no improvement in breast cancer screening performance and increased radiologist interpretation time. No prior study had evaluated whether tracking changes in AI case scores over sequential screenings could improve cancer detection and recall decisions for digital breast tomosynthesis (DBT).

The institution deployed ProFound AI (version 2.0, iCAD) on four Hologic Selenia Dimensions DBT units, generating per-examination case scores (0–100) for 31,741 screenings. A retrospective analysis of 1,799 patients with two or more sequential screenings assessed whether the direction and magnitude of score change over time correlated with true-positive, false-positive, true-negative, and false-negative outcomes.

True-positive examinations had the highest average case score (75) and the largest mean score increase (+21.1) between sequential screenings, both statistically significant versus true-negative (P < .001) and false-positive (P = .02) groups. Combining a case score threshold of ≥26 with a score-change threshold of ≥+1 achieved 93.6% sensitivity and 62.8% specificity, outperforming either threshold alone. The authors conclude that tracking score trajectory alongside absolute score can meaningfully support radiologist recall decisions.

• Tracking AI score change over sequential screenings adds diagnostic value beyond a single absolute score, particularly for identifying true-positive cancers early.
• A combined threshold (score ≥26 AND change ≥+1) meaningfully improves sensitivity (93.6%) while maintaining useful specificity (62.8%), reducing both missed cancers and unnecessary recalls.
• Persistently high scores without large changes warrant careful lesion-level review, as benign findings (e.g., vascular calcifications) can confound the case-level score.