Program Information
Mask-Free Three-Dimensional Digital Subtraction Angiography (3D-DSA) Using a Convolutional Neural Networks-Based Deep-Learning Method
JC Montoya*, Y Li , C Strother , GH Chen , University of Wisconsin-Madison, Madison, WI
Presentations
MO-AB-FS4-7 (Monday, July 31, 2017) 7:30 AM - 9:30 AM Room: Four Seasons 4
Purpose: The purpose of this work was to develop a deep-learning method, based on convolutional neural networks (CNN), to generate 3D-DSA angiograms from a single contrast-enhanced exam without mask acquisition.
Methods: Clinical image volumes of 18 patients scanned with standard 3D-DSA imaging protocol for the assessment of cerebrovascular abnormalities were retrospectively collected. Images from 10 subjects were used to extract more than 100 million image patches that were used as training data set. A CNN-based deep-learning method was trained to classify three tissue types (vasculature, bone and soft tissue). The trained CNN deep-learning model was then applied for the task of tissue classification in a test cohort consisting of the remaining image volumes from 8 subjects. The final vasculature tissue class was used to generate the 3D-DSA images. To quantify the generalization error of the trained model, tissue classification accuracy was measured in 5.7 million image patches from clinically relevant regions of interest in the test data sets. Finally, the generated 3D-DSA images were subject to a qualitative assessment for the presence of inter-sweep motion artifacts.
Results: Overall accuracy for tissue classification in the testing dataset was 98.8%. No residual signal from osseous structures was observed for all cases generated using the proposed CNN-based deep-learning method (except for small regions in the mandible and dental implants, far from the anatomy of interest) compared to 62% (5/8) of the cases that presented moderate to severe mis-registration artifacts using vendor’s DSA algorithm.
Conclusion: A CNN-based deep-learning method was developed to generate 3D-DSA images without mask data acquisition. Since there is no mask acquisition involved, the proposed method successfully eliminates mis-registration artifacts induced by inter-sweep patient motion, potentially reduces radiation dose, and improves image quality in clinical 3D-DSA imaging. Further validation is required to generalize our results for patients with complex cerebrovascular abnormalities.
Funding Support, Disclosures, and Conflict of Interest: Dr Chen received funding support from Siemens Healthcare. Dr Strother received funding support from both Siemens Healthcare and Department of Radiology University of Wisconsin Madison.
Contact Email: