Risk-Stratified Use of Concurrent Computer-Aided Diagnosis (CAD) in Chest CT: Gains in Overall Sensitivity with Loss for CAD-Negative Nodules.

This study quantifies how concurrent computer-aided detection/diagnosis (CAD) alters radiologists' performance in chest CT, emphasizing CAD-negative nodules and the trade-off between overall sensitivity and detection of unmarked lesions. We enrolled 150 consecutive patients undergoing thin-section chest CT. A reference standard of 334 nodules (93% < 10 mm) was established by consensus. Six readers (3 novice, 3 expert) interpreted each scan without CAD and, after a 4-week washout, reinterpreted them with CAD overlays visible from the outset (concurrent reading, outside the recommended workflow). We analyzed sensitivity, case-based specificity, and reading time. Overall sensitivity improved from 47 to 59% (p < 0.01) and mean reading time decreased by 42.8 s (p < 0.01). Conversely, sensitivity for the 75 CAD-negative nodules declined significantly (novices, - 20 percentage points; experts, - 16 points; both p < 0.01). Mixed-effects logistic regression confirmed lower odds of detecting CAD-negative nodules with concurrent CAD (OR, 0.34; 95% CI, 0.25-0.47). Gains were driven by 3-9 mm nodules, while < 3 mm lesions showed no benefit. Reader case-based specificity rose from 75 to 82% with CAD assistance. Concurrent CAD improved overall sensitivity and interpretation speed but significantly reduced sensitivity for CAD-negative nodules, consistent with reduced search vigilance for unprompted regions. Because most CAD-negative nodules were ≤ 6 mm, risk-stratified CAD workflows should be validated prospectively before broad clinical adoption.