Abstract
We developed and tested a deep-learning-based algorithm for the approximation of contrast agent dosage based on computed tomography (CT) scout images. We prospectively enrolled 817 patients undergoing clinically indicated CT imaging, predominantly of the thorax and/or abdomen. Patient weight was collected by study staff prior to the examination 1) with a weight scale and 2) as self-reported. Based on the scout images, we developed an EfficientNet convolutional neural network pipeline to estimate the optimal contrast agent dose based on patient weight and provide a browser-based user interface as a versatile open-source tool to account for different contrast agent compounds. Our pipeline predicted patient weight with a mean absolute error of 3.90 ± 0.20 kg in 5-fold cross-validation. Interpretability analysis revealed that both larger anatomical shape and higher overall attenuation were predictive of body weight. Our open-source deep learning pipeline allows for the automatic estimation of accurate contrast agent dosing based on scout images in routine CT imaging studies, streamlining workflows and enhancing patient safety.
The Challenge of Accurate Weight Estimation
In Computed Tomography (CT), contrast agents are essential for visualizing blood vessels and organs. To achieve optimal image quality while minimizing patient exposure, the dose must be precisely tailored to the patient’s body weight. Currently, clinicians rely on self-reported weights or manual measurements. Both methods are problematic: manual weighing adds significant time to clinical schedules, while self-reporting is prone to bias.
As shown in Figure 1, self-reported weights are statistically significantly lower than actual measurements, with an average underestimation of nearly 2 kg. This "vanity gap" can lead to suboptimal dosing if not corrected.
AI-Driven Weight Prediction from CT Scout Images
To address these issues, researchers developed an end-to-end pipeline centered on EfficientNet-B1, a convolutional neural network (CNN) optimized for efficiency and accuracy. The model utilizes the CT "scout" image—a low-dose preview scan captured at the start of every examination—to estimate weight automatically.
Key technical features of the approach include:
- Multi-Region Training: The model was trained on thoracic, abdominal, and combined thoracic-abdominal scans, making it versatile for various clinical scenarios.
- Interpretability: Using Variational Autoencoders (VAEs) and dataset distillation, the team found that the AI considers both anatomical shape and X-ray attenuation (density) to estimate mass.
- Privacy-First Deployment: The tool is implemented as a browser-based interface using ONNX Runtime, allowing it to run locally on hospital hardware without patient data leaving the network.
Performance and Clinical Impact
The study evaluated the model on a prospective cohort of 817 adult patients (832 total scans). The pipeline achieved a Mean Absolute Error (MAE) of 3.90 ± 0.20 kg across the entire dataset (Figure 2).
In clinical terms, this weight error translates to a contrast agent dose difference of only 4.5 to 11.7 ml, which is considered a very narrow and acceptable margin in routine practice. The model performed particularly well on combined thoracic-abdominal scans (MAE of 3.07 kg), likely because a larger field of view provides more information about total body mass. Notably, the model exhibited a slight "regression toward the mean," tending to underestimate very high weights and overestimate very low weights.
Limitations and Future Directions
While the results are promising, several factors must be addressed before widespread adoption. The model's accuracy can be sensitive to patient positioning or poor image quality. Furthermore, as a single-center study, multi-center validation is necessary to ensure the AI generalizes across different scanner manufacturers and patient populations. Currently, the tool is intended to assist clinicians, maintaining a "human-in-the-loop" to verify AI-generated suggestions.
Summary updated