Training of a deep learning based digital subtraction angiography method using synthetic data.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Cardiovascular Imaging

Background:

• Digital Subtraction Angiography (DSA) is crucial for diagnosing cardiovascular diseases (CVDs).
• Deep learning-based DSA (DDSA) offers dose reduction and improved image quality but is limited by artifact-prone clinical data.
• Current DDSA models struggle with specific structures and noise-free image prediction due to data limitations.

Purpose of the Study:

• To develop a strategy for generating abundant synthetic DSA image pairs for DDSA model training.
• To create synthetic DSA targets free from artifacts and noise common in clinical DSA.

Main Methods:

• Utilized over 7,000 CT projection images and 25,000 synthetic vascular projection images to create synthetic DSA image pairs.
• Generated vessel skeletons using stochastic Lindenmayer systems and CT scans.
• Trained DDSA models on the synthetic dataset and compared performance against models trained on clinical DSA data.

Main Results:

• DDSA models trained on synthetic data performed comparably or better than those trained on clinical data across leg, abdomen, and cardiac imaging.
• Synthetic data training resulted in clearer DSA-like images, outperforming conventional DSA and clinical data-trained models.
• Quantitative evaluations showed superior peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), accuracy, precision, and Dice scores for models trained on synthetic data.

Conclusions:

• A novel approach for training DDSA networks using synthetic DSA image pairs was proposed.
• This method enables direct extraction of DSA-like images from contrast-enhanced x-ray images.
• The developed approach shows potential as a valuable tool for aiding in cardiovascular diagnosis.