Document resource
Introduction/Purpose Catheter Digital Subtraction Angiography (DSA) is the gold standard for visualizing blood vessels during endovascular interventions, but it is heavily degraded by voluntary, respiratory, or cardiac motion. This motion creates artifacts that can obscure vascular detail, often requiring repeat acquisitions that increase procedural time, radiation, and risk. To address this unmet clinical need, we developed NeuroClear, a fully automated deep learning algorithm for motion-resistant background subtraction neuroangiography.Materials and Methods Three distinct deep learning approaches were developed to overcome motion degradation using spatiotemporal input. The most successful method utilized a database of >8,000 deformable motion patterns combined with a motion augmentation pipeline to generate synthetic datasets. These datasets were used to train a neural network to perform background subtraction utilizing deformable registration. The algorithm was evaluated on a multi-site, prospective test set of consecutive neuroangiographic studies in awake patients. Three expert clinicians graded the algorithm against standard-of-care DSA using a 5-point Likert scale (1=markedly inferior, 3=similar, 5=markedly superior) evaluating Sharpness, Noise, Artifact, Vascular Fidelity, and Overall Quality. Real-time feasibility was evaluated by implementing the algorithm on NVIDIA Clara AGX edge hardware.Results For Overall Quality, the NeuroClear algorithm never underperformed traditional DSA, meeting the noninferiority design goal >95% of the time. It performed similarly to DSA in 4.6% of cases (which featured very little motion), outperformed DSA in 52.9% of cases, and markedly outperformed DSA in 42.5% of cases. The deep learning algorithm achieved high mean Likert scores for Artifact reduction (4.56), Vascular Fidelity (4.27), and Overall Quality (4.38), significantly surpassing DSA across metrics (p<0.005). Furthermore, at a full native resolution of 1024x1024, the average inference time was ~60 milliseconds per frame. This speed easily satisfies the requirements for real-time neuroangiography, which typically operates at 2 to 6 frames per second.Conclusion Motion-resistant artificial background subtraction with deep learning is highly robust and markedly outperforms traditional DSA in vascular fidelity and overall image quality. Its successful implementation on edge hardware yields sub-100ms real-time inference, demonstrating its viability for immediate, intra-procedural clinical application to improve the safety and efficiency of neurointerventional procedures.Disclosures D. Cantrell: 1; C; NIH/NHLBI, AHA, Nvidia. 4; C; Clearvoya, LLC. A. Shaibani: None. C. Cantrell: 4; C; Clearvoya, LLC. S. Ansari: 1; C; NIH/NHLBI, AHA, Nvidia. 4; C; Clearvoya, LLC.Abstract O-035 Figure 1Sample images comparing DSA to our NeuroClear deformable registration algorithm trained with synthetic data, demonstrating marked improvement in image quality and vascular fidelity