Deep learning-based frame synthesis enables radiation dose reduction in digital subtraction angiography imaging: a...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cerebrovascular Diseases

Background:

• Digital subtraction angiography (DSA) is crucial for diagnosing and treating cerebrovascular diseases.
• Low-dose imaging in DSA is vital to minimize radiation risks for patients and healthcare professionals.
• Current DSA protocols necessitate balancing diagnostic needs with radiation exposure concerns.

Purpose of the Study:

• To develop and validate SAVE-Net, a deep learning model for synthesizing DSA frames.
• To reduce the number of scans required in clinical practice, thereby lowering radiation dose.
• To maintain diagnostic image quality while minimizing radiation exposure in cerebrovascular imaging.

Main Methods:

• SAVE-Net was trained and validated on 17,335 DSA sequences from one hospital and externally validated on 3,255 sequences from two additional hospitals.
• The model synthesizes intermediate DSA frames using deep learning and optical flow estimation.
• Quantitative metrics (SSIM, PSNR) and a visual Turing test by interventional radiologists assessed image quality and diagnostic performance.

Main Results:

• SAVE-Net generated DSA sequences with high consistency to real clinical data, achieving SSIM of 0.951 and PSNR of 40.764 in external validation.
• Generated sequences showed no significant difference in image quality and diagnostic confidence compared to real DSA data.
• The model achieved this using only 1/7 of the standard radiation dose, with a generation time of 0.04 s/frame.

Conclusions:

• SAVE-Net effectively generates clinically diagnostic DSA sequences with significantly reduced radiation exposure (1/7th dose).
• The synthesized images maintain comparable image quality and diagnostic confidence to standard DSA.
• This AI-driven approach offers a practical solution for reducing radiation risks in cerebrovascular imaging.