Automated scout-image-based estimation of contrast agent dosing: a deep learning approach.
Authors
Affiliations (5)
Affiliations (5)
- Division of Medical Physics, Department of Diagnostic and Interventional Radiology, Medical Center - University of Freiburg, Faculty of Medicine - University of Freiburg, Freiburg, Germany.
- Department of Diagnostic and Interventional Radiology, Medical Center - University of Freiburg, Faculty of Medicine - University of Freiburg, Freiburg, Germany.
- Department of Stereotactic and Functional Neurosurgery, Medical Center - University of Freiburg, Faculty of Medicine - University of Freiburg, Freiburg, Germany.
- Department of Diagnostic and Interventional Radiology, Medical Center - University of Freiburg, Faculty of Medicine - University of Freiburg, Freiburg, Germany. [email protected].
- Department of Neuroradiology, Medical Center - University of Freiburg, Faculty of Medicine - University of Freiburg, Freiburg, Germany.
Abstract
Optimal contrast agent dosing in computed tomography (CT) depends on accurate patient weight, yet manual measurements increase workload and self-reporting can introduce bias. We developed and tested a deep-learning-based algorithm to automate the approximation of contrast agent dosage directly from CT scout images. We retrospectively analyzed 817 patients undergoing thorax/abdomen CT. Prior to examination, patient weight was collected via manual scale measurements and self-reporting. We developed an EfficientNet convolutional neural network pipeline to estimate weight from scout images and used in-context learning and dataset distillation to analyze body-weight-informative CT features. The model was used in a browser-based user interface to provide dosing estimates for various contrast agent compounds. Self-reported patient weights were statistically significantly lower than manual scale measurements (75.13 kg vs. 77.06 kg; p < 10<sup>- 5</sup>, Wilcoxon signed-rank test). In 5-fold cross-validation, the pipeline predicted patient weight with a mean absolute error (MAE) of 3.90 ± 0.20 kg. This error corresponds to a difference of roughly 4.48-11.70 ml of contrast agent, depending on the specific agent. Interpretability analysis confirmed that both larger anatomical shape and higher overall attenuation were the predictive features of body weight. This open-source deep learning pipeline enables automatic, accurate contrast agent dosing in routine CT workflows. The approach has the potential to improve patient safety and clinical efficiency by providing accurate weight estimates without requiring additional measurements or relying on potentially outdated records. Further validation on larger cohorts and across different clinical centers is required.