Reducing motion artifacts in craniocervical background subtraction angiography with deformable registration and unsupervised deep learning.

Background: In clinical practice, digital subtraction angiography (DSA) often suffers from misregistration artifact resulting from voluntary, respiratory, and cardiac motion during acquisition. Most prior efforts to register the background DSA mask to subsequent postcontrast images rely on key point registration using iterative optimization, which has limited real-time application.

Purpose: Leveraging state-of-the-art, unsupervised deep learning, we aim to develop a fast, deformable registration model to substantially reduce DSA misregistration in craniocervical angiography without compromising spatial resolution or introducing new artifacts.

Materials and methods: We extend HyperMorph, an open source deep learning deformable registration framework, to reduce motion artifacts in DSA. Novel image similarity loss functions with vessel layer estimation were introduced to optimize background registration, making it robust to the variable presence of intravascular iodinated contrast.

Results: A total of 516 studies with 5,240 angiographic series were collected and divided into training (5,046 series) and hold-out test (194 series) sets. Blinded algorithm rankings and Likert scores on 5-point scales (1 = worst, 5 = best) were generated by 3 practicing interventional neuroradiologists using 50 series randomly selected from the hold-out test set. Compared to traditional DSA, our learning-based background subtraction angiography (BSA) significant improved vascular fidelity (2.4 ± 0.6 for DSA vs. 3.6 ± 0.5 for BSA), subtraction artifacts (2.0 ± 0.4 for DSA vs. 3.9 ± 0.3 for BSA), and overall quality (2.1 ± 0.5 for DSA vs. 3.9 ± 0.4 for BSA) (P < .0001). Learning-based BSA also significantly outperformed affine registration-based BSA (P < .0001). The average inference time for learning-based BSA was 30 milliseconds per frame on our hardware.

Conclusion: The results demonstrate that deep learning deformable registration, combined with an appropriate loss function, can significantly reduce the motion artifacts that degrade DSA.