To compare the diagnostic performance of radiology residents and attending radiologists in detecting retained surgical items (RSIs) on intraoperative radiographs.

In this pilot study, 100 de-identified intraoperative radiographic cases (18 positive for RSIs, 82 negative) were reviewed. For each case, upper-level radiology resident (PGY-3 to PGY-5) and attending radiologist participants recorded the presence or absence of an RSI, their confidence on a three-point scale, and their decision time. We compared accuracy, sensitivity, specificity, confidence, and interpretation time between the two groups. We fit a multivariable logistic regression (fixed-effects GLM) to identify predictors of a correct interpretation, followed by a mixed-effects logistic regression (GLMM) with random intercepts for reader and case to account for clustering.

A total of 1,178 interpretations were analyzed (619 from residents, 559 from attendings). There was no significant difference in diagnostic accuracy between residents (94.2%) and attendings (94.3%) (Fisher’s exact p = 0.84 for accuracy, 0.82 for sensitivity, and 0.74 for specificity). Attendings were slightly faster (median time 16.4s vs. 18.8s; p = 0.0038) and reported higher confidence (mean 2.65 vs. 2.52; p < 0.001). In the fixed-effects GLM, participant type was not associated with accuracy (resident vs. attending AOR 0.92, 95% CI 0.52–1.59, p = 0.76). Compared with ‘unsure,’ being ‘somewhat confident’ (AOR 9.75, 95% CI 4.72–20.4) and ‘very confident’ (AOR 20.9, 95% CI 9.71–46.4) markedly increased the odds of a correct interpretation (both p < 0.001). Longer response times were associated with lower odds of correctness (AOR 0.66, 95% CI 0.46–0.94, p = 0.020). Aside from non-significant associations with response time and foreign-body type, findings were otherwise consistent in the mixed-effects model.

Upper-level radiology residents demonstrate diagnostic accuracy for detecting retained surgical items that is statistically indistinguishable from that of attending radiologists. These preliminary findings suggest that a resident-led preliminary interpretation model for RSI studies at the point of service is a feasible and potentially efficient approach that would not compromise patient safety.