Coronary Artery Calcium Identification on Non-Gated Sim CT Using a Convolutional Neural Network
Abstract
Purpose
To develop a convolutional neural network (CNN) to detect coronary artery calcium (CAC) on non-gated (free heartbeat and free breathing) simulation CT scans and evaluate the robustness of models trained on different datasets, including gated (ECG-gated with breath hold) and non-gated CT scans.
Methods
Training, validation, and testing data were obtained from multiple institutions (Stanford, WashU) and clinical imaging cohorts (Chest CT, Cardiac CT, breast cancer) and scored using established methods. Scans with CAC scores in the range [1-10] were excluded from training due to lack of clinical significance. All models used the 3D Densenet121 CNN architecture. The CT images were masked and cropped to the heart using the heart contour, and a threshold of 100-700 HU was applied to isolate CAC. We investigated how gated vs. non-gated acquisitions can affect the robustness and generalization of the CAC model. First, a model was trained using only gated data (N=376), then multiple models were trained gradually incorporating more non-gated data (N=376 to 494). All models were tested and compared on a test set of 262 CT images (80 gated, 182 non-gated, excluding CAC scores [1-10]).
Results
The model trained on only gated data achieved 90% accuracy on the test set with FNR (false negative rate) of 0.09 and FPR (false positive rate) of 0.11. The model trained on gated and 100% non-gated data achieved 87% accuracy on the same test set with FNR of 0.10 and FPR of 0.17.
Conclusion
A CNN was trained to detect CAC on both gated and non-gated CT scans with high accuracy. Models trained using gated data demonstrated comparable performance to those trained with mixed gated and non-gated data, suggesting that appropriate preprocessing and thresholding can enable robust CAC detection on non-gated CT scans without including non-gated data during training.