The Double Edged Impact of AI-Assisted PE Detection on Junior and Senior Radiology Residents

Purpose:
To evaluate the impact of an Artificial Intelligence (AI) decision-support tool on diagnostic performance and decision-making among junior and senior radiology residents interpreting CT examinations for pulmonary embolism (PE).

Materials and Methods:
In this prospective multi-reader study, 10 radiology residents (5 junior, pre-Board examination; 5 senior, post-Board examination) independently interpreted 145 CT examinations, including dedicated CT pulmonary angiography, TAVI, and routine contrast-enhanced CT protocols. The cohort comprised 113 AI-concordant cases and 32 AI-erroneous cases (15 false negatives, 17 false positives). Each reader performed a standardized three-step evaluation: (1) independent interpretation, (2) review of AI output (Aidoc), and (3) final diagnostic decision. Diagnostic accuracy, sensitivity, specificity, and change-of-mind rates were assessed against an expert thoracic radiologist reference standard.

Results:
A total of 1,450 interpretations were analyzed. Junior residents were significantly more likely to change their initial diagnosis following AI review than senior residents (9.9% vs. 3.4%; OR=3.12, p<0.001). AI assistance produced a substantial net diagnostic benefit in junior residents (+3.03% accuracy), with sensitivity increasing from 84.3% to 93.4% (p<0.001). However, juniors also demonstrated greater susceptibility to automation bias, with more harmful changes from correct to incorrect diagnoses (3.45% vs. 1.66%; OR=2.12, p=0.038) and a lower rate of correct overrides of erroneous AI outputs (72.8% vs. 86.7%, p=0.013). Among senior residents, AI increased sensitivity from 92.3% to 95.4% (p=0.001), but resulted in only a minimal net diagnostic benefit (+0.14%). Both groups demonstrated a slight reduction in specificity following AI exposure.

Conclusion:
AI-based PE detection has a double-edged impact on radiology residents. While it substantially improves junior residents’ sensitivity and reduces the diagnostic gap with senior residents, it simultaneously increases their vulnerability to erroneous AI recommendations. Senior residents demonstrate greater resilience to AI errors, suggesting that the diagnostic benefit of AI is strongly dependent on reader experience.