Calculation of virtual 3D subtraction angiographies using conditional generative adversarial networks (cGANs)

Area of Science:

• Medical imaging
• Artificial intelligence in radiology
• Vascular imaging

Background:

• Subtraction angiography requires two scans, increasing patient radiation dose.
• Bone and metal artifacts can obscure vessels in standard angiography.
• Generative AI offers potential to simulate subtraction angiography, reducing radiation exposure.

Purpose of the Study:

• To evaluate the efficacy of conditional generative adversarial networks (cGANs) in simulating 3D subtraction angiographies.
• To determine if AI-generated angiographies maintain image quality and diagnostic accuracy while reducing radiation.
• To assess artifact reduction and potential new artifact introduction by the AI model.

Main Methods:

• A conditional generative adversarial network (cGAN) model was trained using 621 3D subtraction angiographies.
• The model was adapted from the pix2pix framework for 3D data.
• AI-generated angiographies were compared against original datasets by five blinded neuroradiologists, assessing similarity, quality, and artifacts.

Main Results:

• AI-simulated subtraction angiographies demonstrated no significant difference in image quality or diagnostic accuracy compared to original scans.
• Bone and movement artifacts were reduced in the AI-generated images.
• Artifacts from metal implants varied, with no significant superiority of either method.

Conclusions:

• Conditional generative adversarial networks can effectively simulate subtraction angiographies for clinical use.
• This AI technology has the potential to reduce patient radiation dose.
• The simulation process may introduce novel artifacts that require further investigation.