Reference-Free, Learning-Based Rigid Motion Compensation in Cone-Beam CT for Interventional Neuroradiology
H Huang1*, J Siewerdsen1,2, W Zbijewski1, C Weiss2, T Ehtiati3, A Sisniega1, (1) The Johns Hopkins University, Baltimore, MD (2) Russell H. Morgan Department of Radiology, Johns Hopkins University, Baltimore, MD (3) Siemens Healthineers, Forchheim, Germany
Presentations
TU-C-TRACK 3-7 (Tuesday, 7/27/2021) 1:00 PM - 2:00 PM [Eastern Time (GMT-4)]
Purpose: Cone-beam CT (CBCT) provides guidance in interventional neuroradiology, but limited acquisition speed (4 – 30 s) results in motion artifacts. Autofocus motion compensation (MoCo) was successfully applied in head CBCT but autofocus metrics do not guarantee preservation of anatomical structures. Similarity metrics (e.g., visual information fidelity, VIF) provide quantification of image quality and structural similarity to a reference. However, motion-free, matched, references are seldom available. We propose DL-VIF, a learned reference-free autofocus metric that preserves anatomy.
Methods: DL-VIF was estimated with a deep CNN implementing a cascade of 3D residual blocks acting on downsampled head CBCTs (128³ voxels). The network was trained on simulated motion-free and motion-contaminated CBCT pairs. Random rigid motion trajectories (0-8mm amplitude) were applied to 38 head MDCT volumes (12100 instances). DL-VIF was integrated into an autofocus framework and assessed on 20 CBCTs generated analogously to the training data. Further validation was obtained on 5 CBCTs of a cadaver head acquired on a robotic C-arm (Artis Zeego, Siemens Heathineers) with motion induced by shifting and rotating the specimen. The performance of DL-VIF was compared to a conventional autofocus metric (gradient entropy).
Results: DL-VIF showed good agreement with the reference (slope = 0.980, R² = 0.996). DL-VIF MoCo showed reduction of motion artifacts in simulated data, with 0.099 improvement in SSIM, compared to 0.094 for conventional autofocus. In the cadaver study, DL-VIF outperformed conventional autofocus, providing robust compensation in cases of moderate and severe motion, where conventional autofocus failed. An overall 6x larger increase in SSIM and 7x reduction in SSIM standard deviation was observed, restoring visualization of cranial anatomy.
Conclusion: The learned DL-VIF provided reliable estimation of structural similarity without a reference image and yielded consistent reduction of motion artifacts with anatomically realistic solutions. DL-VIF offers an important step in robust autofocus MoCo methods for CBCT.
Funding Support, Disclosures, and Conflict of Interest: Academic-industry collaboration with Siemens Healthineers.
Handouts
Keywords
Cone-beam CT, Motion Artifacts, Reconstruction
Taxonomy
IM- Cone Beam CT: Development (New Technology and Techniques)
Contact Email