Room: AAPM ePoster Library
Purpose: To predict contrast enhancement in transcatheter aortic valve replacement (TAVR) CT using machine learning with input patient information and acquisition parameters.
Methods: A neural network with two distinct input and output layers was created using the TensorFlow framework. Retrospective data containing known patient and acquisition parameters from 116 TAVR CT exams was split into training (85.5%), validation (4.5%), and testing (10%) datasets. Patient mass, patient height, and thoracic acquisition kVp were fed into the first input layer to estimate thoracic aortic enhancement at the first output layer. Predicted enhancement from the first output layer was concatenated with abdominal acquisition kVp and injected contrast volume at the second input layer to predict abdominal aortic enhancement at the second output layer. Thoracic and abdominal aortic enhancement were measured as the mean HU from ROIs manually drawn near the aortic root or in the abdominal aorta on axial TAVR images. Mean squared error (MSE) and mean absolute error (MAE) between predicted and measured enhancement were used as the training loss metric and overall performance metric, respectively.
Results: Using the best-trained weights, the neural network predicted thoracic and abdominal aortic enhancement in the test dataset with MAE values of 34 HU and 44 HU and standard deviations of 43 HU and 57 HU, respectively. Differences in predicted and measured enhancement in the thoracic and abdominal aorta varied by greater than 50 HU in 16.7% and 41.7% of cases, respectively, for patient BMI values ranging from 26–49.
Conclusion: Machine learning based on known patient information and acquisition parameters demonstrates promise for predicting contrast enhancement in TAVR CT using retrospective data. It is expected that a well-trained neural network could prospectively predict whether or not enhancement in TAVR CT images would be adequate, and identify high-contrast exams where reducing contrast dose may be appropriate.