Introduction Creating 3D-DSA scans of the brain results in notable radiation exposure for patients and staff.¹ Nevertheless, the quality of the subtraction series is frequently undermined by patient motion.

Aim of Study This study aims to generate subtraction volumes exclusively from contrast-enhanced series using a Generative Adversarial Network (GAN), eliminating the need for native images.

Methods A customized GAN, inspired by pix2pix,² was trained on 779 3D-DSA datasets, split 80:20 for training and validation. Evaluation used two test datasets: 72 with motion and metal artifacts (Data_wA) and 92 without major artifacts (Data_woA). All scans were performed with a Siemens Axiom Artis. Experienced neuroradiologists performed the rating using a custom-built app.

Results The results showed high-quality generated series based on PSNR, SSIM and NMSE (figure 1c). The pairwise delta distribution (figure 1d) indicated that the generated series were rated as better, equal to, or only one grade worse than the original series, with improved ratings for motion artifacts.

• Download figure
• Open in new tab
• Download powerpoint

Abstract O34 Figure 1

shows (a) the conceptual schematic; (b) 3D contrast agent, subtraction (with motion artifact) and the generated GAN image; (c) the PSNR, SSIM and NMSE results for both test datasets; (d) the evaluation delta for DatawA (positive values mean better evaluation of the generated)

Conclusion The GAN effectively emphasized blood vessels without additional information like vessel segmentation masks and the diagnostic accuracy of the generated series was not inferior compared to the real 3D subtraction series. It reduced motion artifacts but showed limitations by metal implants without significant superiority of one group. Addressing these requires expanding training data and refining GAN parameters.

Synthetic 3D-DSA images without native references are achievable with a GAN, mitigating motion artifacts and reducing radiation exposure. So, the GAN can be an alternative to threshold-based methods.

References

1. Bor Dogan, et al. ‘Patient and staff doses in interventional neuroradiology.’ Radiation Protection Dosimetry 2005;117(1-3):62-68.

2. Choi Jae Won, et al. ‘Generating synthetic contrast enhancement from non-contrast chest computed tomography using a generative adversarial network.’ Scientific Reports 2021;11(1):1-10.

Disclosure of Interest no.