Patient-informed modelling of hepatic contrast dynamics in contrast-enhanced CT imaging

Abstract

PURPOSE Iodinated contrast agents are commonly used in CT imaging to enhance tissue contrast. Consistency in contrast enhancement (CE) is critical in radiological diagnosis. Contrast material circulation in individual patients is affected by factors such as patient body habitus and anatomy leading to significant variability in organ contrast enhancement, image quality, and dose. Toward the goal of improving CE consistency in clinical populations, in this work we developed a contrast dynamics model to predict CT HU enhancement of liver parenchyma in abdominopelvic CE CT scans.

METHOD AND MATERIALS This study included 700 adult abdominopelvic contrast CT exams performed in 2014-2018 using two scanner models from two vendors. Each CT image was segmented using a deep learning-based segmentation algorithm and the hepatic parenchyma HU values were acquired from the segmentations. A two-layer neural network-based algorithm was used to identify the relationship between patient attributes (height, weight, BMI, age, sex), scan parameters (slice thickness, scanner model), contrast injection protocols (bolus volume, injection-to-scan wait time), and the liver HU CE. We randomly selected 60% studies for training, 10% validation, and 30% for testing the accuracy. The training output was the extracted HU values. The goodness-of-fit of the model was evaluated in terms of R^2, Adjusted R^2, Mean Absolute Error (MAE), and Mean Squared Error (MSE) between the model prediction and ground truth. In addition, the generalizability of the model was evaluated by comparing the R^2 in the training data (leave-one-out validation) and the testing data.

RESULTS This preliminary model has an 0.51 R^2, 0.40 adjusted R^2, 10.0 HU MAE, 159.1 HU MSE, 0.6±12.8 HU Mean Error, and 2.5 HU Median Error on test data. For training data, the model has 0.59 R^2, 0.56 Adjusted R^2, and 0.5 predicted R^2. The close R^2 between testing and training data results indicate a reasonable generalizability.

CONCLUSION Results showed considerable predictability of liver CE from patient attributes, scanning parameters, and contrast administration protocol. We envision to expand the model to include other major organs toward a comprehensive predictive model.

CLINICAL RELEVANCE/APPLICATION A contrast dynamics model can be an essential tool to personalize contrast-enhanced CT protocol and to improve the consistency of contrast enhancement across different patients in diagnostics imaging.