Quantitative assessment of video frame interpolation in coronary angiography using a novel spatio-temporal blur metric.

Background

Synthetic fluoroscopy frames interpolated between real frames have been proposed as a method of reducing radiation doses in coronary angiography. Traditional full‑reference image quality assessment (FR‑IQA) metrics often favour strongly smoothed images and correlate poorly with expert cardiologist assessment of clinical image quality.

Methods

Retrospective coronary angiography data from 29 patients (58 sequences at 7.5 and 15 frames per second) were processed using linear interpolation (LI), Residue Refinement Interpolation (RRIN) and Real‑time Intermediate Flow Estimation (RIFE) algorithms. A three‑reader observer study, based on DIMOND coronary angiography quality criteria, provided clinical image‑quality scores for each sequence. A Global Noise Index (GNI), adapted from computed tomography, and BM3D, a novel extension of an established spatial blur metric into the spatio-temporal domain, were calculated for all datasets.

Results

Original datasets achieved the highest observer scores for resolution and branch visibility, with RRIN significantly lower and RIFE not significantly different from originals. GNI ranked algorithms RRIN < LI ≈ RIFE < Original, indicating that RRIN achieved the greatest noise suppression but was rated inferior due to over‑smoothing. BM3D ranked algorithms Original < RIFE < RRIN < LI (lower values indicating less blur), matching the clinical rankings and showing higher blur at 7.5 fps than at 15 fps.

Conclusions

BM3D provides a transparent spatio‑temporal blur metric that aligns with expert assessment of coronary angiography image quality, in contrast to traditional FR‑IQA metrics. BM3D is therefore a promising tool for objective assessment and optimisation of clinical dynamic x-ray imaging.