Impact of Exposure Parameters on Deep Learning Models in Chest Radiography and Implications for Deployment.
Authors
Affiliations (5)
Affiliations (5)
- Department of Computer Science, National Tsing Hua University, Hsinchu, Taiwan.
- Laboratory for Computational Physiology, Massachusetts Institute of Technology, Cambridge, Mass.
- Division of Pulmonary Critical Care and Sleep Medicine, Beth Israel Deaconess Medical Center, Boston, Mass.
- Department of Radiology, Emory University, Atlanta, Ga.
- Department of Biostatistics, Harvard T.H. Chan School of Public Health, Boston, Mass.
Abstract
Purpose To investigate whether deep learning models trained on chest radiographs (CXRs) 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, CXRs from MIMIC-CXR (January 2011-December 2016), the Medical Imaging and Data Resource Center (MIDRC; August 2020-May 2022), and EmoryCXR (September 2008-February 2023) were analyzed for pneumothorax detection, coronavirus disease 2019 (COVID-19) diagnosis, and race classification. Dataset-provided labels served as the reference standard. Three exposure parameters (ExposureTime, XRayTubeCurrent, ExposureInuAs) were extracted from Digital Imaging and Communications in Medicine (DICOM) 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. Area under the receiver operating characteristic curve (AUC) was compared using the DeLong test. Results A total of 727,604 CXRs from 240,681 patients (mean age, 60 years ± 17 [SD]; 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 (ΔAUC = -0.38; <i>P</i> < .001). Similar declines were observed for COVID-19 (ΔAUC = -0.33; <i>P</i> < .001) and race classification (ΔAUC = -0.09; <i>P</i> < .001). The 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 CXRs may exploit exposure parameters as shortcut features; exposure-regimen audits may flag high-risk conditions before clinical deployment. ©RSNA, 2026.