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

Background: As an important imaging tool for diagnosing and treating cerebrovascular diseases, the low-dose imaging technology of digital subtraction angiography (DSA) can effectively reduce radiation exposure risks for both patients and operators. To ensure clinical demand while minimizing radiation dose, this study proposed SAVE-Net, which integrates deep learning with optical flow estimation. The model is designed to synthesize intermediate frames in DSA sequences, thereby reducing the number of scans required in clinical practice and directly decreasing the radiation dose. Methods: SAVE-Net was developed to generate subsequent frames following any given real DSA frame. A total of 17,335 DSA sequences from one hospital were used for model training, fine-tuning, and internal validation. For external validation, an additional 3,255 DSA sequences from two other hospitals were utilized. Image similarity between generated and real frames was quantitatively evaluated using the Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio (PSNR). Furthermore, five interventional radiologists independently performed a visual Turing test and quality assessment on the generated sequences. Inter-rater agreement for the Turing test results was assessed using Fleiss' Kappa, while the Wilcoxon Signed-Rank test was employed to analyze significant differences in the quality ratings. Results: Internal validation results demonstrated that SAVE-Net achieved DSA sequences highly consistent with real clinical data using only 1/7 of the standard radiation dose and maintained stable performance across multiple scenarios. External validation results illustrated that SAVE-Net achieved an effective performance [SSIM: 0.951, 95% CI: (0.948, 0.956); PSNR: 40.764, 95% CI: (40.673, 40.798); and generation time: 0.04 s/frame]. Assessment results showed no significant difference between the generated sequences and the real ones. Additionally, the generated results exhibited high consistency with real data in terms of overall image quality (4.919 vs. 4.940) and diagnostic confidence (4.838 vs. 4.910). Conclusion: SAVE-Net enables the generation of clinically diagnostic DSA sequences using only 1/7 of the standard radiation dose, with image quality and diagnostic confidence comparable to real clinical data. Its superior performance across multi-center validation demonstrates a practical and effective approach to reducing radiation exposure in cerebrovascular imaging.