Background Subtraction Angiography with Deep Learning Using Multi-frame Spatiotemporal Angiographic Input.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Catheter Digital Subtraction Angiography (DSA) is susceptible to motion artifacts from patient movement during image acquisition.
• Previous machine learning approaches for DSA image enhancement focused on individual 2D frames, neglecting temporal information.
• Motion artifacts degrade the diagnostic quality of DSA, hindering accurate vessel visualization.

Purpose of the Study:

• To develop and evaluate improved 2D+t deep learning models for motion artifact reduction in DSA.
• To leverage temporal information from angiographic time series for enhanced image quality.
• To introduce a synthetic motion augmentation pipeline for training deep learning models.

Main Methods:

• Collected 516 cerebral angiograms (8784 series) and classified them into 'motionless' and 'motion-degraded' subsets.
• Developed a synthetic motion augmentation pipeline using feature-based computer vision algorithms.
• Trained and evaluated 2D U-Net, 3D U-Net, SegResNet, and UNETR models on the augmented dataset.

Main Results:

• The 3D U-Net model significantly outperformed 2D U-Net architectures in reducing motion artifacts.
• 3D U-Net achieved a lower RMSE (23.14 ± 9.56) and higher Multi-Scale SSIM (0.93 ± 0.05) compared to single-frame 2D U-Net.
• 3D U-Net demonstrated competitive performance against other convolutional and transformer-based models like SegResNet and UNETR.

Conclusions:

• Incorporating multi-frame temporal information substantially improves the performance of motion-resistant deep learning algorithms for DSA.
• The developed synthetic motion augmentation pipeline is effective for training 3D (2D+t) deep learning architectures.
• The 3D U-Net model shows significant potential for reducing motion artifacts in neuroangiography, leading to improved image clarity.