摘要
<p>Purpose: To investigate whether deep learning models trained on chest radiographs rely on radiographic exposure parameters as shortcut features and to quantify the resulting biases under controlled confounding and natural exposure regimes. Materials and Methods: In this retrospective study, chest radiographs from MIMIC-CXR (January 2011-December 2016), the Medical Imaging and Data Resource Center (August 2020-May 2022), and EmoryCXR (September 2008-February 2023) were analyzed for pneumothorax detection, COVID-19 diagnosis, and race classification. Dataset-provided labels served as the reference standard. Three exposure parameters (ExposureTime, XRayTubeCurrent, and ExposureInuAs) were extracted from Digital Imaging and Communications in Medicine metadata. Models were trained under biased and balanced exposure label alignments and evaluated on matched and reversed distributions. A priori screening additionally identified high-risk exposure regimes. The area under the receiver operating characteristic curve (AUC) was compared using the DeLong test. Results: A total of 727 604 chest radiographs in 240 681 patients (mean age +/- SD, 60 years +/- 17; 126 432 men, 114 128 women) were analyzed. For pneumothorax detection, AUC decreased from 0.94 (95% CI: 0.94, 0.95) to 0.56 (95% CI: 0.55, 0.58) on mismatched exposure distributions (Delta AUC, -0.38; P < .001). Similar declines were observed for COVID-19 (Delta AUC, -0.33; P < .001) and race classification (Delta AUC, -0.09; P < .001). The a priori exposure-regimen screening revealed high-risk regimes within the natural distribution that were associated with reduced model performance compared with typical exposures. Conclusion: Deep learning models trained on chest radiographs may exploit exposure parameters as shortcut features; exposure-regimen audits may flag high-risk conditions before clinical deployment.</p>