Novel Use of an AI-Based Video Interpolation Algorithm to Reduce Radiation Dose in the Cardiac Catheterization Laboratory.

BACKGROUND: Radiation dose reduction is a target for innovation in cardiology. Frame rate reduction during image acquisition in coronary angiography reduces radiation dose but at the cost of clinical information. Video frame interpolation (VFI) is an artificial intelligence (AI)-based technique which can synthesize intermediate frames between existing ones and may be suited to fluoroscopic imaging. VFI algorithms, therefore, represent an opportunity to reduce radiation doses while maintaining the same frame rate. AIMS: Our aim was to evaluate two VFI algorithms, Real-time Intermediate Flow Estimation (RIFE) and Residue Refinement Interpolation (RRIN), in their ability to reduce radiation doses in the catheterization laboratory. METHODS: Sixty coronary angiography sequences were retrospectively obtained. Every second frame from each sequence was replaced with an artificial frame synthesized by VFI algorithms (RIFE and RRIN). Each altered sequence was compared to the original quantitatively using computer vision metrics. Sequences were assessed qualitatively by blinded cardiologists using a novel scoring system based on clinically relevant image quality criteria. Lin's concordance correlation coefficient between scores for each algorithm and original sequences was then calculated. RESULTS: On clinical assessment, RIFE slightly outperformed RRIN. RIFE showed good correlation with original sequences for noise suppression (0.747, 95% CI 0.682-0.812), main branch evaluation (0.737, 95% CI 0.672-0.802), and side branch evaluation (0.792, 95% CI 0.740-0.844). Computer vision metrics indicated that RRIN performed better than RIFE. Results from qualitative and quantitative assessments disagreed on the best-performing VFI algorithm. CONCLUSION: VFI through AI-based algorithms may allow radiation dose reduction while maintaining diagnostic capability. The superior clinical results and low training and computational footprint of the RIFE algorithm make it attractive for real-time frame rate augmentation.