Video frame interpolation for coronary angiography using latent flow matching

Abstract

Radiation exposure is inherent to coronary angiography. Lowering its frame rate significantly reduces radiation exposure but degrades temporal resolution. We introduce Angio-FILM, a novel generative AI model for video frame interpolation dedicated to coronary angiography. Angio-FILM uses a latent flow matching model to synthesize high-temporal-resolution (15 FPS) videos from low-frame-rate (7.5 FPS) inputs. The model was trained on 357,933 videos and validated on both internal and external open-source datasets. While quantitative metrics showed mixed results, qualitative evaluation by human experts showed Angio-FILM outperformed existing state-of-the-art methods. Furthermore, Visual Turing test (30 physicians, 600 videos) showed accuracies of 54% (p = 0.107) in a single-video binary classification task and 49% (p = 0.749) in a two-alternative forced choice task, indicating that the generated videos were hardly distinguishable from real angiograms. Quantitative coronary analysis confirmed high anatomical fidelity, showing minimal deviation in minimal lumen diameter measurements between the original and interpolated frames (MAE, 0.180 mm). Angio-FILM represents a significant step forward in translating a cutting-edge generative AI technology into a clinically helpful tool.

Introduction

Coronary angiography, the gold standard for evaluating the coronary arteries, is an X-ray-based imaging technique¹. Due to the dynamic nature of the heart and rapid coronary blood flow, it is typically taken in a multi-frame video sequence at 10–15 frames per second (FPS)^2,3. While ionizing radiation enables the acquisition of these images, it also poses significant hazards to both patients and operators, including risks of cancer, cataract, and skin injury⁴.

Efforts have been made to reduce radiation dose during coronary angiography and interventions⁵. Because the cumulative dose is proportional to the total number of frames exposed, reducing the frame rates of fluoroscopy and cine acquisition has been one of the representative strategies to reduce radiation exposure. Established clinical consensus on radiation safety underlines the direct relationship between frame rate and radiation dose, advising practitioners to employ the lowest dose rate that does not compromise image quality^6,7. Multiple studies have shown that lowering frame rates to 7.5 or 10 FPS can significantly reduce radiation exposure without compromising image quality, procedural and patient outcomes^8–12. However, in clinical practice, physicians often prefer higher frame rates to avoid perceived image degradation, such as flicker and jerkiness^12,13.

Video frame interpolation (VFI) is a classic computer vision problem that aims to synthesize intermediate frames between two consecutive video frames¹⁴. It has been applied to various purposes, including slow-motion generation, frame rate up-conversion, video compression, novel view synthesis, video restoration, and intra-prediction in video coding for VR games or smartphone cameras. Its methodology has evolved over decades, including optical flow-based, kernel-based and phase-based methods. Recently, advances in generative artificial intelligence (AI) have remarkably upgraded its performance^15,16.

In this study, we introduce Angio-FILM, a novel generative AI-based VFI model dedicated to coronary angiography (Fig. 1). While several studies have applied VFI technology to the medical domain^17,18, coronary angiography poses specific challenges, such as gray-scale images where optical flow estimation is difficult, the nature of projection images where multiple structures are overlapped, large motion of coronary arteries along the cardiac cycle, and various motion artifacts originating from the patient’s movement and respiration. As illustrated in Fig. 2, existing VFI methods often struggle with these inherent complexities, whereas Angio-FILM effectively captures precise vessel structures even under severe motion. Therefore, the primary objective of this study was to develop a high-quality, high-fidelity VFI model capable of overcoming these challenges. Developing such a model is the first step toward the clinical goal of reducing radiation exposure without compromising diagnostic quality.

Fig. 1. Angio-FILM model architecture and study workflow.

a Schematic showing velocity field learning to generate 15 frames per second (FPS) frames from 7.5 FPS inputs. b Workflow overview including model development on 377,002 videos and validation via Visual Turing tests and anatomical analysis.

Fig. 2. Qualitative comparison of VFI models under large cardiac motion.

Consecutive frames i to i + 6 demonstrate motion continuity across seven columns. Rows compare ground truth against RIFE, LDMVFI, LBBDM, and Angio-FILM. Blue boxes highlight interpolated frames; 2 by 2 grid lines serve as spatial references.

The main contributions are summarized as follows: (1) An extended-context generative framework for periodic medical dynamics: We move beyond traditional local-interpolation and stochastic diffusion models by introducing Angio-FILM, specifically designed for the non-linear, pulsatile dynamics of coronary angiography. By combining latent flow matching with an extended (N + 1) temporal conditioning principle, the model leverages the periodic cardiac cycle as a global constraint, ensuring geometric consistency and structural integrity even under severe motion. (2) Functional decoupling to mitigate clinical hallucinations: In medical imaging, the risk of a generative model “inventing” a lesion (hallucination) is critically high. We strictly decouple spatial reconstruction (the autoencoder) from temporal synthesis (the denoising transformer) to ensure structural fidelity. The autoencoder is mathematically constrained to the data manifold of real images, ensuring that the transformer can only synthesize intermediate frames that are spatially and diagnostically valid. (3) Rigorous clinical validation of generative fidelity: We demonstrate that standard computer vision metrics are inadequate for coronary angiography and propose a more rigorous clinical evaluation standard. By validating our framework through a large-scale Visual Turing test (30 physicians) and quantitative coronary angiography (QCA), we establish that the generated frames preserve critical vascular geometry and are virtually indistinguishable from real recordings.

Results

Overview of Angio-FILM

Angio-FILM is a flow-matching-based VFI model in which (N + 1) temporally consecutive frames are provided as conditioning inputs to generate N intermediate frames. Flow matching is a generative learning approach that estimates velocity fields to transform a source distribution into a target distribution¹⁹. In this framework, an intermediate frame I_n between two known neighboring frames, the past frame I₀ and future frame I_1, is transformed from Gaussian noise with guidance of (N + 1) neighboring frames. We empirically set N = 15, while an alternative value was evaluated in the ablation study (see “Ablation study”). Although training was performed with a fixed N, inference can be performed with variable N by interpolating the positional embeddings. Following the latent diffusion model (LDM) paradigm²⁰, we employed the Cosmos-Tokenizer²¹, a pretrained autoencoder, to enhance memory and computational efficiency. Among the autoencoders evaluated on the CAG dataset, Cosmos-Tokenizer demonstrated superior performance, achieving the best scores across PSNR, MS-SSIM, and FID, thereby ensuring exceptional reconstruction fidelity (Supplementary Table 1); it was therefore selected as our primary model (see “Choice of autoencoder” in “Methods”) (see Supplementary Videos 1–3 for frequently observed artifacts related to autoencoders). Conditioning frames were first mapped into a latent space via the encoder, where Angio-FILM performed the interpolation. The generated latent representations of the target frames were then reconstructed back into the image space through the decoder (Supplementary Fig. 1). The denoising network was implemented as a Transformer (Latte²²), offering greater scalability and cross-domain adaptability than conventional U-Net designs.

The model was trained using 357,933 videos derived from 36,098 studies across 27,461 patients collected from Seoul National University Bundang Hospital (SNUBH) (Table 1). This internal dataset encompasses a wide clinical spectrum, including normal coronary angiograms, stenotic lesions, chronic total occlusions (CTOs), peripheral angiograms with contrast, and sequences acquired during stent implantation. The dataset was split into training, validation, and test sets in a 95:2.5:2.5 ratio. External validation was conducted on a public dataset, CoronaryDominance, which contains 10,543 videos from 1574 patients²³. The best-performing model, identified through quantitative metrics, expert evaluation, and ablation study, was subsequently tested in a Visual Turing Test by human experts.

Table 1.

Baseline characteristics of the study population

Characteristics	All (N = 28,907)	Training set (N = 27,461)	Validation set (N = 723)	Test set (N = 723)
Age (years)	66.0 ± 12.3 (26,370)	66.0 ± 12.3 (25,039)	65.6 ± 12.0 (656)	66.8 ± 12.1 (675)
Sex (male)	17,535/26,396 (66.4%)	16,657/25,064 (66.5%)	451/656 (68.8%)	427/676 (63.2%)
BMI (kg/m²)	24.7 ± 3.5 (24,334)	24.7 ± 3.5 (23,117)	25.0 ± 3.4 (595)	24.5 ± 3.5 (622)
Current smoker	4847/18,652 (26.0%)	4600/17,728 (25.9%)	122/460 (26.5%)	125/464 (26.9%)
Hypertension	13,820/26,447 (52.3%)	13,102/25,115 (52.2%)	350/656 (53.4%)	368/676 (54.4%)
Diabetes	7824/26,447 (29.6%)	7408/25,115 (29.5%)	199/656 (30.3%)	217/676 (32.1%)
Dyslipidemia	5990/26,447 (22.6%)	5674/25,115 (22.6%)	153/656 (23.3%)	163/676 (24.1%)
History of stroke	1404/26,447 (5.3%)	1330/25,115 (5.3%)	28/656 (4.3%)	46/676 (6.8%)
Chronic kidney disease	870/26,447 (3.3%)	835/25,115 (3.3%)	20/656 (3.0%)	15/676 (2.2%)
Clinical diagnosis
Stable coronary disease	14,667/22,606 (64.9%)	13,905/21,460 (64.8%)	396/596 (66.4%)	366/550 (66.5%)
Unstable Angina	2632/22,606 (11.6%)	2517/21,460 (11.7%)	52/550 (9.4%)	63/596 (10.6%)
NSTEMI*	2976/22,606 (13.2%)	2807/21,460 (13.1%)	87/550 (15.8%)	82/596 (13.8%)
STEMI*	2331/22,606 (10.3%)	2231/21,460 (10.4%)	45/550 (8.2%)	55/596 (9.2%)
LV* ejection fraction (%)	57.9 ± 14.6 (19,665)	57.9 ± 13.9 (18,658)	58.4 ± 31.6 (496)	58.0 ± 11.6 (511)

*NSTEMI denotes non-ST elevation my ocardial infarction, STEMI ST-elevation myocardial infarction, LV left ventricular

Quantitative evaluation of Angio-FILM against baseline methods

Table 2 summarizes the interpolation performance of Angio-FILM and three baseline VFI models, LDMVFI¹⁶, Consecutive Brownian Bridge Diffusion (LBBDM)²⁴, and RIFE²⁵, trained on the same SNUBH split for a fair comparison. RIFE, an optical flow-based approach, synthesizes intermediate frames via flow-field estimation, whereas LDMVFI and LBBDM are diffusion-based frameworks that generate intermediate frames through iterative denoising or Brownian bridge diffusion processes. Evaluations were performed on both internal and external coronary angiography datasets using PSNR, MS-SSIM, LPIPS, and FID with DINOv2 features²⁶. PSNR and MS-SSIM assess pixel-level fidelity and structural similarity (higher is better), whereas LPIPS captures perceptual similarity in deep feature space (lower better). The FID, which is computed using DINOv2 features²⁷ to mitigate the texture bias inherent in Inception-V3 and achieve a more shape-aware evaluation of distributional realism, measures the perceptual distance between the generated and real frames (lower better).

Table 2.

Quantitative comparison of Angio-FILM with different video interpolation SOTA methods

	PSNR (↑)	MS-SSIM (↑)	LPIPS (↓)	FID (↓)
Internal
RIFE	34.6441	0.9392	0.3854	162.0180
LDMVFI	33.2646	0.9232	0.0948	8.1819
LBBDM	33.3938	0.9260	0.0960	6.6189
Angio-FILM (ours)	34.3207	0.9429	0.1365	3.0989
External
RIFE	31.7319	0.9293	0.4323	122.1502
LDMVFI	30.2931	0.9125	0.1075	12.9106
LBBDM	30.5535	0.9178	0.1165	13.2714
Angio-FILM (ours)	31.0683	0.9263	0.1827	5.4750

Internal dataset indicates coronary angiography videos from SNUBH, and the external dataset corresponds to the CoronaryDominance public dataset. Angio-FILM was evaluated with 50 sampling steps, while LDMVFI used DDIM with 200 steps and LBBDM used 50 steps. The best performance for each metric is highlighted in bold

Across both the internal and external datasets, Angio-FILM achieved the best FID scores, while RIFE obtained the highest PSNR values, and LDMVFI achieved the lowest LPIPS. Angio-FILM and RIFE each performed best in terms of MS-SSIM on the internal and external datasets, respectively. Comparison of additional quantitative metrics is also presented in Supplementary Table 2. However, manual examination of each model’s outputs revealed that these quantitative metrics often failed to align with perceptual quality (Fig. 3). Despite its favorable PSNR and MS-SSIM, RIFE produced markedly blurred videos that lacked clear vascular contours—a feature essential for identifying coronary lesions. This inconsistency highlights the limitation of PSNR and MS-SSIM in coronary angiography: they tend to assign high scores to overly smooth frames that obscure diagnostically important details. Similarly, LPIPS frequently misjudges failure cases, such as overlaid or distorted coronary structures, by assigning falsely low (i.e., erroneously favorable) scores. In addition, when evaluating the autoencoder’s reconstruction quality, we observed that LPIPS exhibited notable instability, assigning substantially different scores to visually indistinguishable frames (Supplementary Fig. 2; see “Choice of autoencoder” in “Methods”). Collectively, the observations led us to conclude that conventional metrics are inadequate to assess clinically meaningful fidelity in this domain. Therefore, we proceeded with a formal human expert evaluation as the next stage.

Fig. 3. Discrepancies between quantitative metrics and clinical fidelity.

Examples show metrics failing to reflect perceptual quality. a, b RIFE achieves higher PSNR and MS-SSIM despite blurred vessels. c, d LPIPS erroneously favors baseline outputs despite vessel and catheter distortion. FPS denotes frames per second.

Expert evaluation on interpolation quality across models

A formal human expert evaluation was deemed necessary because conventional metrics yielded mixed results and aligned poorly with visual assessment. For this, four interventional cardiologists compared Angio-FILM and baseline VFI methods in a two-stage blinded evaluation. Prior to testing, the experts reviewed representative failure cases and discussed visual cues for identifying interpolation artifacts (Fig. 4 and Supplementary Video 4). The evaluation time was unlimited, and both video clips and frame-level still images were provided, allowing them to freely play or pause the videos for frame-level detailed inspections.

Fig. 4. Characteristic interpolation artifacts and failure cases.

Panels illustrate common errors identified by experts : a overlapping structures, b blurring, c lesion distortion, d small vessel distortion, e catheter distortion, and f device deformation. Supplementary Video 4 details dynamic artifacts.

In the first stage, paired videos generated using the four different study models were compared against the original one: four randomly arranged interpolated versions were parallelly presented along with the original video, and evaluators were asked to select the one most similar to the original (Supplementary Video 5). Fifty images were provided for each evaluator. The closest model to the ground truth assessed by human evaluators was shown to be Angio-FILM (92%) >> LBBDM (7%) > LDMVFI (1%) > RIFE (0%) (Supplementary Table 3).

The quality of individual videos was evaluated in the second stage; the cinematic realism and photorealism of each output were rated using a 10-point quality scale by cardiologists. In total, 150 videos from five different models (one original and four interpolated) on the internal and external datasets were provided in a random and blinded manner. The estimated mean score was original ≥ Angio-FILM >> LBBDM > LDMVFI > RIFE with high inter-observer agreement (intraclass correlation coefficient: 0.93 [0.91–0.95]) (Fig. 5). While the quantitative metrics showed mixed results among the models, expert evaluation clearly demonstrated Angio-FILM outperformed other baseline models across internal and external datasets.

Fig. 5. Human expert evaluation of video realism across datasets.

Mean 10-point realism scores for internal (a) and external (b) data. Blue bars indicate marginal mean scores; error bars represent 95% confidence intervals from a linear mixed model. Dashed lines represent original video scores.

Visual Turing test: real vs. fake by human experts

Thirty clinicians with varying clinical experience participated in the visual Turing test, which consisted of two stages: a binary classification task followed by a two-alternative forced-choice (2AFC) task (Fig. 6). In both stages, the raters were given unlimited viewing time and were able to play or pause each video for frame-by-frame inspection. The first stage was a binary classification task on 600 videos (300 real + 300 interpolated) randomly sampled from the test cohort. If physicians perform at a chance level of 50% accuracy, the model is considered to have “passed” the test by being indistinguishable from the real one. The overall accuracy, adjusted for inter-rater variability, was 54% (95% CI: 49%–58%), which was not significantly different from the null hypothesis of 50% (p = 0.113). The accuracy did not show significant variation according to the physicians’ clinical experience. The sensitivity and specificity were 56% (50–61%) and 52% (46–57), respectively (Supplementary Table 4). If raters identified a video as AI-generated, they were asked to assess the image quality (score of 1–10). Among 288 responses on 133 real and 155 interpolated videos, the score was 7.73 ± 0.20 and 7.71 ± 0.20 (p = 0.837).

Fig. 6. Visual Turing test results for clinician discriminability.

a, d Web-based evaluation interfaces. b, e Clinician accuracy across experience levels for binary and two-alternative forced choice tasks. Blue squares and horizontal lines represent mean accuracy and 95% confidence intervals. c Quality scores for videos judged as AI-generated.

The second stage was conducted in a 2AFC design, where 300 pairs of videos (one real and one interpolated) were provided side by side and the participating physicians were asked to determine which one was real. The overall accuracy adjusted for inter-rater variability was 49%, which was not significantly different from the chance level of 50% (p = 0.749). Consistent results were observed across all rater groups, confirming that the interpolated videos were virtually indistinguishable by human experts.

Quantitative validation: anatomical fidelity via QCA

To quantitatively validate the anatomical fidelity of the proposed model, we conducted a quantitative coronary angiography (QCA) study in a blinded manner. Detailed procedures for the QCA standardization and measurement protocols are described in the “Methods” section (see “QCA Standardization and Measurement Protocol” in “Methods”). The deviation from the original images and the interpolated ones was estimated using MAE and RMSE. The correlation between the AI-QCA and manual QCA measurements was represented by a scatter plot and evaluated using Pearson’s correlation coefficient. The Bland–Altman plot was used to visualize the agreement between the two methods.

The quantitative comparison demonstrated a high degree of agreement between the MLD measurements from the original and interpolated frames (Fig. 7). The mean absolute error (MAE) and root mean square error (RMSE) were 0.180 mm and 0.261 mm, respectively, indicating minimal deviation from the ground truth. The 95th percentile of absolute error was 0.576 mm, further confirming the model’s reliability in preserving vascular dimensions even in worst-case scenarios. Statistical analysis showed a strong positive correlation between the two groups, with a Pearson’s correlation coefficient (R) of 0.879 (95% CI: 0.857–0.898). Furthermore, Bland-Altman analysis revealed a mean bias of –0.056 mm with limits of agreement (LoA) ranging from –0.555 to 0.443 mm. These results suggest that the proposed VFI model maintains high anatomical fidelity and provides clinically acceptable diagnostic accuracy, effectively addressing concerns regarding generative hallucinations in pathological regions.

Fig. 7. Quantitative MLD comparison between original and interpolated frames.

a Correlation analysis of measurements showing the linear regression as a solid line. b Bland–Altman plot illustrating mean bias as a blue line and 95% limits of agreement as red dashed lines. MLD denotes minimal lumen diameter.

Ablation study

We conducted ablations to validate the design choices of Angio-FILM. Following our observation that quantitative metrics do not always align with the perceptual fidelity required for clinical angiography, our final configurations were selected based on a combination of metric trends, qualitative stability, and internal expert review. A detailed comparison of the performance metrics across different model configurations is presented in Table 3.

Table 3.

Quantitative ablation on training configurations and inference strategies of Angio-FILM

Method	PSNR(↑)	MS-SSIM(↑)	LPIPS(↓)	FID (↓)
Input sequence length
N = 1	33.4133	0.9282	0.1517	4.9036
N = 3	33.9492	0.9365	0.1411	3.6897
N = 7	34.1092	0.9391	0.1359	2.4304
N = 15 (*)	34.3207	0.9429	0.1365	3.0989
Denoising network scale
Latte L/4	34.2324	0.9413	0.1353	2.9459
Latte XL/4 (*)	34.3207	0.9429	0.1365	3.0989
CFG masking unit
Sample-level masking	-	-	-	-
Frame-level masking (*)	34.3207	0.9429	0.1365	3.0989
Conditioning strategy
Cross-attention	30.4201	0.9092	0.1972	45.0048
In-context (*)	34.3207	0.9429	0.1365	3.0989
ECG-guided sampling
w/ ECG	34.3233	0.9429	0.1364	3.1090
w/o ECG (*)	34.3207	0.9429	0.1365	3.0989
Number of sampling steps
5 (cfg 1.5)	34.2647	0.9499	0.1488	12.8943
10 (cfg 2.0)	34.5771	0.9467	0.1352	3.6932
15 (cfg 2.5)	34.4737	0.9452	0.1355	2.9321
25 (cfg 2.5)	34.4117	0.9440	0.1352	2.6762
50 (cfg 3.0)	34.3207	0.9429	0.1365	3.0989
75 (cfg 3.0)	34.2953	0.9425	0.1366	3.1735
100 (cfg 3.0)	34.2803	0.9423	0.1368	3.1860
200 (cfg 3.0)	34.2562	0.9420	0.1371	2.3863

The first four categories (Input sequence length, Denoising network scale, CFG masking unit, and Conditioning strategy) evaluate design choices during the training phase. The remaining two categories focus on the inference-time strategies of sampling steps and ECG guidance, conducted using the optimized model. Asterisks (*) denote the final model configuration selected for Angio-FILM. The best results for each metric in each section are highlighted in bold

The appearance of the coronary arteries on angiography varies substantially throughout the cardiac cycle. The accuracy and fidelity of Angio-FILM depend on its ability to capture and reflect these complex dynamics. To better model this, Angio-FILM was designed to take (N + 1) consecutive frames as input and generate N intermediate frames, rather than generating a single intermediate frame by referencing the preceding and following frames. This design allows the model to leverage temporal context across a broader range of cardiac cycles. To determine the appropriate context window size, we conducted an ablation study in which the values varied.

The model trained with N = 15 achieved the highest PSNR and MS-SSIM values, whereas the model trained with N = 7 yielded the best LPIPS and FID scores (Table 3). As noted earlier, PSNR and MS-SSIM can be misleading if they reward blurry outputs (Fig. 3). However, in diffusion-based (e.g., LDMVFI, LBBDM) and flow-matching–based models such as Angio-FILM, we observed no noticeable blurring artifacts, making them more reliable for relative comparison in this context. Given that our task provides a clear ground truth for each interpolated frame, we therefore prioritized full-reference metrics (PSNR and MS-SSIM) over distributional ones (FID).

To complement these quantitative results with a perceptual evaluation, we conducted a visual inspection of several large-motion samples (Supplementary Fig. 3). We observed that the N = 1 was notably less stable than N > 1 frame, although it still demonstrated greater robustness to large motion compared to baseline models that also interpolate a single frame. Among the N = 3, 7, and 15 variants, no discernible differences were found in terms of the structural and motion accuracy of the main vessels.

Ultimately, N = 15 was selected as the final configuration. This choice was informed by its superior PSNR and MS-SSIM performance and also aligns with clinical physiological constraints. Assuming a standard frame rate of 7.5 FPS and a contrast agent duration of approximately 2 s, N = 15 ensures that the model leverages the maximum amount of available information—specifically, the full set of frames where contrast is visible. While a smaller N might be more appropriate for videos with lower frame rates, N = 15 was finalized to ensure broader applicability and robustness in standard clinical settings.

We investigated the impact of the denoising network scale by comparing Latte L/4 and Latte XL/4. While PSNR and MS-SSIM favored Latte XL/4, LPIPS and FID favored Latte L/4 (Table 3). To identify potential perceptual nuances that standard metrics might fail to capture, we performed a qualitative assessment on a limited set of large-motion samples, specifically looking for vessel discontinuity, blurring, or overlapping artifacts. Despite subtle visual differences between the two scales, we selected Latte XL/4, following the original Latte study’s finding that larger models yield significant performance gains even on small-scale datasets.

We investigated different strategies for masking conditioning frames during classifier-free guidance (CFG) training. Specifically, we compared sample-level masking—where the entire sequence of (N + 1) conditioning frames is masked for a subset of samples in a batch—against frame-level masking, in which each conditioning frame within a sample is masked independently based on a set probability. Under identical training conditions, we observed that sample-level masking led to significant training instability, with the loss function failing to reach convergence. Consequently, frame-level masking was adopted to ensure stable training and robust model convergence.

We investigated the influence of different integration mechanisms for the (N + 1) conditioning frames. Specifically, we compared the in-context approach—which involves concatenating target and conditioning frames in their natural temporal order—against a cross-attention mechanism inspired by PixArt-α²⁸. Following 500 k training steps for both configurations, the in-context method consistently outperformed the cross-attention approach across all evaluation metrics (Table 3). Visual observations indicated that the cross-attention model failed to reach adequate convergence within the same training budget, specifically, as demonstrated in Supplementary Video 6. The generated samples exhibited persistent flickering artifacts due to inconsistent stylistic rendering, even though the general vascular structure was captured. This confirms that the in-context strategy provides a more efficient and effective inductive bias for capturing the specific temporal dynamics of coronary angiography.

ECG provides unique information included in coronary angiography, which dictates myocardial and coronary dynamics along the cardiac cycle. While our model was trained with ECG information to leverage these dynamics, we hypothesized that the temporal cues within the angiogram sequence itself might be sufficient during inference. To investigate this hypothesis, we conducted an ablation study to compare model performance with and without ECG guidance during sampling.

As shown in the lower section of Table 3, ECG-guided sampling yielded marginal improvements in PSNR, MS-SSIM, and LPIPS while showing no notable change in the FID score. This minimal performance gap suggests that multiple frames of an angiogram sequence itself provide sufficient temporal and physiological cues for effective frame interpolation. Therefore, we selected the sampling strategy without ECG guidance for our final version. This independence from auxiliary signals is a significant advantage of Angio-FILM, particularly in clinical scenarios in which ECG data may not be available, underscoring the broader applicability of Angio-FILM with robustness.

Sampling steps refer to the number of iterations the generative model takes to refine an image from random noise into a final output. To investigate the effect of the number of sampling steps on the performance of Angio-FILM, we conducted an ablation study using our best-performing model (Table 3 and Supplementary Fig. 4).

Although we used 50 sampling steps for the model evaluation, we found that a comparable performance was achievable with fewer steps. In our tests, the performance peaked at different stages: MS-SSIM at 5 steps, PSNR at 10, LPIPS at 10 and 25, and FID at 25. Using only 10 steps, for instance, would reduce the total sampling time by approximately 80% while maintaining interpolation quality. Nevertheless, even at 50 steps, Angio-FILM remained more efficient than or comparable to baseline models such as LDMVFI (DDIM 200 steps) and LBBDM (50 steps).

Inference latency benchmark

To evaluate the clinical feasibility of our model, we measured the inference latency of Angio-FILM and the baseline models under a simulated clinical scenario. In clinical practice, VFI is primarily required only for the segment where the radiocontrast agent is present. Based on our test dataset, the average duration of contrast-filled frames was 43, with 90% of the videos having a contrast duration of 62 frames or less. Therefore, we established a benchmark task of converting a 31-frame sequence acquired at 7.5 FPS into a 15 FPS video, which requires generating 30 intermediate frames. All latency measurements were conducted on an NVIDIA H100 GPU with an Intel Xeon Gold 6430 CPU. For the baseline models, which synthesize a single intermediate frame between two consecutive input frames, we measured the latency using a batch size of 2. In contrast, Angio-FILM conditions on 16 input frames to simultaneously generate 15 intermediate frames, allowing us to evaluate it with a batch size of 1. The optical flow-based model, RIFE, achieved the fastest inference time at 0.101 ± 0.000 s. Among the generative models, LBBDM and LDMVFI recorded latencies of 16.851 ± 0.139 s and 63.405 ± 0.474 s, respectively. Angio-FILM completed the same task in 17.256 ± 0.010 s with 50 sampling steps. This demonstrates that Angio-FILM achieves competitive latency while offering enhanced structural efficiency through simultaneous multi-frame synthesis.

Discussion

In this study, we propose Angio-FILM, which generates coronary angiography videos of high fidelity, doubling the frame rate from 7.5 to 15 FPS. To our knowledge, this is the first approach that applies VFI to coronary angiography as a conditional generation problem using latent flow matching. The model architecture was carefully designed, integrating a pre-trained autoencoder, frame conditioning with classifier-free guidance, and a denoising network. The model was trained and validated on a huge body of coronary angiography videos. Angio-FILM was comprehensively evaluated and demonstrated superiority over previous SOTA models on internal and external datasets using both quantitative metrics and subjective evaluation. Finally, the output of the model was virtually indistinguishable from real recordings by clinicians, including experienced interventional cardiologists, who assessed the image quality of the real and generated videos that were highly similar.

Coronary angiography presents unique challenges for VFI, as described in the Introduction. Our study confirmed that previously identified issues, such as artifacts from large motion and occlusion, remain significant hurdles for previous VFI methods^14,16,24,25. We observed that these interpolation artifacts were more prevalent in the right coronary artery (RCA) and left circumflex (LCX) artery than in the left anterior descending (LAD) artery. This appears because RCA and LCX are situated near the base of the left ventricle, a region that undergoes substantially greater motion throughout the cardiac cycle than the cardiac apex, where distal LAD is located²⁹. In a line, these motion artifacts were most prominent in views where such movement is exaggerated, such as the right anterior oblique (RAO) view for RCA and the cranial views for LCX.

Beyond these previously reported challenges, our study highlights two issues specific to coronary angiography: inadequate rendering of small vessels and unnatural blood flow. The coronary arterial system consists of large epicardial arteries and a network of smaller vessels³⁰, the details of which have been poorly captured by previous methods. Experienced cardiologists readily identified this lack of detail as a notable artifact even when the quantitative metrics did not differ significantly.

Another major issue noted by clinicians was the unrealistic rendering of the fluid dynamics. Angiography visualizes radiocontrast filling, which is a direct consequence of the underlying blood flow. This flow is not a continuous stream, but is uniquely pulsatile, reflecting the heart’s contractions and relaxations. Conventional VFI methods, however, interpolated the radiocontrast movement in a linear fashion—akin to “water flowing”—failing to capture the complex, rhythmic dynamics inherent to the cardiac cycle (Supplementary Video 4).

The field of VFI has been transformed by the advent of generative AI, which moves beyond traditional pixel blending to synthesize novel photorealistic content^15,16. A key finding of our study is that generative models significantly outperformed the optical flow-based model, RIFE. This result highlights a fundamental limitation of the traditional VFI methods in this domain. While optical flow-based models operate by estimating pixel motion and warping existing frames, generative models learn to synthesize entirely new photorealistic content based on data distribution. This advantage allows generative models to create plausible intermediate frames in challenging scenarios in the medical field, such as the complex nonlinear dynamics of coronary angiography.

In this study, we also learned that our model, based on flow matching, worked best among other generative models, such as diffusion-based LDMVFI and LBBDM. Diffusion model and flow matching are two prominent generative architectures up to date^19,31. While diffusion models excel at creating diverse, high-quality outputs, their stochastic denoising process risks generating plausible but anatomically incorrect details—a potential for “clinical hallucination” that is unacceptable for diagnostic or procedural use. Therefore, we hypothesize that the deterministic trajectory approach is better suited for coronary angiography. Flow matching models learn a direct and continuous transformation between frames, which encourages smooth and geometry-preserving mappings similar to optimal transport flows³². The results of this study validate our hypothesis: Angio-FILM not only outperformed other models but also produced videos that clinicians found indistinguishable from real recordings, demonstrating that a deterministic approach like flow matching is highly suitable for this clinical application.

Angio-FILM holds considerable promise for clinical applications, primarily due to its potential for direct integration into existing coronary angiography and intervention workflows. The primary clinical benefit would be the reduction in radiation exposure for both patients and physicians, which could be achieved by acquiring angiograms at a lower frame rate. Further improvements are required to realize its full potential in live procedures. Currently, processing a 6-s angiogram acquired at 7.5 FPS to 15 FPS takes approximately 25 s with 50 sampling steps, and about 5 s when the number of steps is reduced to 10, showing comparable performance without notable degradation; latency benchmarked with 50 warm-up iterations and averaged over 100 runs. However, real-time use during procedures requires near-instant inference. Another practical limitation is the requirement of high computing power, which may not be available in all catheterization laboratories. Finally, the clinical value of this approach must be rigorously validated. Future randomized controlled clinical trials are essential to prospectively confirm whether this technology can indeed lower radiation doses without compromising diagnostic accuracy, patient safety, or procedure time.

In conclusion, Angio-FILM is the first VFI model to demonstrate that a latent flow matching approach can generate coronary angiography videos that are clinically realistic and indistinguishable from real recordings. We anticipate that Angio-FILM will facilitate the clinical adoption of lower frame rate protocols, thereby providing a viable option for reducing radiation exposure without compromising diagnostic quality. More broadly, our work represents a significant step forward by successfully translating a cutting-edge generative AI technology into a practical tool with clear clinical benefits

Methods

Data collection and curation

The dataset consisted of two sources: an internal dataset from SNUBH and a public dataset, CoronaryDominance. Model training was performed exclusively on the training split of SNUBH, whereas evaluation was conducted on the test split of SNUBH as well as on CoronaryDominance, allowing us to assess both in-domain and cross-domain generalizations.

Coronary angiography videos acquired at SNUBH between 2003 and 2023 were collected. All videos were recorded at 15 FPS with an average duration of 5.34 s. The cohort spanned a broad clinical spectrum, including normal coronary angiograms, stenotic lesions, chronic total occlusions (CTO), peripheral angiograms with contrast, and images obtained during stent implantation. We restricted the dataset to studies that contained ECG traces that were initially considered useful for conditioning our model. In addition, we included only videos with at least 31 frames, as this length was necessary for training the model to take 16 even-indexed frames as inputs and interpolate the 15 intervening odd-indexed frames. The dataset comprised 28,907 patients, 38,001 studies, and 377,002 videos. We performed a patient-wise split of 95:2.5:2.5 into training, validation, and test sets, yielding 27,461/723/723 patients, 36,098/951/952 studies, and 357,933/9,622/9,447 videos for training, validation, and testing, respectively. This study was approved by the Institutional Review Board of the Seoul National University Bundang Hospital (approval number: B-2507-983-301). Informed consent was obtained from all participants included in the study. All research involving human participants was performed in accordance with the Declaration of Helsinki.

CoronaryDominance contains 1574 studies organized into three parts—main, real distribution, and domain shift—and includes challenging cases, such as poor quality, artifacts, and CTO. All videos were recorded at 15 FPS and did not include the ECG. For our experiments, we selected only those sequences with ≥31 frames to match the temporal window used during inference.

Data preprocessing

To capture complete cardiac motion patterns rather than partial phases, we set the conditioning clip length to $(⌊7.5\times c⌋+1)$ frames, spanning approximately c cardiac cycles under a nominal 60 BPM and 15 FPS assumption. When c = 1, the conditioning clip consists of 8 frames and the model is required to interpolate 7 intermediate (target) frames, resulting in 15 frames per training sample. For c = 2, these numbers were doubled (16 conditioning and 15 target frames), yielding 31 frames per training sample. For each training sample, a contiguous segment of this total length was randomly extracted with even-indexed frames serving as conditioning inputs and odd-indexed frames as interpolation targets.

ECG signals were parsed from the DICOM Curve Data elements recorded during angiography acquisition and processed using the ecg_process from neurokit2³³. We only used sequences for which the ECG could be successfully extracted and processed. For each sampled clip, the ECG segment temporally aligned to the input clip was resampled by interpolation to a fixed length of 100 and supplied as an auxiliary conditioning feature.

Overview of flow matching

Flow matching is a generative modeling framework that formulates data synthesis as integrating an ordinary differential equation (ODE) over time¹⁹. The goal is to learn a continuous transformation x_t that evolves from a simple prior distribution P₀ (e.g., Gaussian noise) to the data distribution P₁. This transformation is governed by a time-dependent velocity field ${\mathrm{v}}_{\mathrm{\theta }}\left({\mathrm{x}}_{\mathrm{t}},\mathrm{t}\right)$ as follows:

$$\frac{d{x}_t}{\mathit{dt}}=v_{\mathrm{\theta }}\left(x_t,t\right),x_0\sim p_0$$ 1

The model thus predicts at each time step the velocity vector that drives the intermediate samples along a designed path toward the data manifold.

Training is typically carried out by constructing a simple reference path between noise and data, often through linear interpolation between a data sample x₁ and a noised sample x₀:

$$x_t=\left(1-t\right)x_0+t{x}_1.$$ 2

Along this path, a target velocity can be analytically defined as the derivative of x_t with respect to t:

$$v^{*}\left(x_t,t\right)=\frac{d{x}_t}{\mathit{dt}}=x_1-x_0.$$ 3

The model is trained to minimize a regression loss between the predicted and target velocities:

$$\mathcal{L}\left(\mathrm{\theta }\right)=E{x}_0,x_1,t\left[|v\mathrm{\theta }\left(x_t,t\right)-v^{*}\left(x_t,t\right){|}^2\right].$$ 4

Once trained, data generation is achieved by numerically solving the learned ODE from t = 0 to t = 1, thereby transporting samples from the prior P₀ to the data distribution P₁. This approach provides a principled and flexible way to model complex data distributions while maintaining a simple training objective and an efficient sampling procedure.

Model architecture

Angio-FILM is a VFI model trained to take (N + 1) consecutive frames as inputs and infer the N temporally intermediate frames that bridge them. By design, the model can accommodate any frame rate because only the positional embeddings must be adjusted. During training, we sampled (2N + 1) consecutive frames from each sequence and provided them to the model; the even-indexed (N + 1) frames were used as the condition, while the odd-indexed N frames were used as the target.

Unlike general natural videos, coronary angiography (CAG) exhibits unique periodic characteristics synchronized with the cardiac cycle. Under the condition that the coronary arteries are sufficiently opacified by a contrast agent, the vessel morphology remains relatively consistent at equivalent phases of the cardiac cycle. Leveraging this domain-specific property, we designed a multi-frame interpolation architecture that utilizes N + 1 conditioning frames. While state-of-the-art video frame interpolation (VFI) models for the natural domain, such as LDMVFI and LBBDM, typically rely on only two adjacent conditioning frames, our model incorporates a broader temporal context. By integrating more than two conditioning frames, Angio-FILM achieves enhanced robustness against large motion, which is often a critical bottleneck in cardiac imaging. This architectural choice ensures that the model captures long-range temporal dependencies, mitigating motion artifacts and maintaining temporal consistency.

To improve the memory and computational efficiency, all frames were encoded into a latent space using a pretrained autoencoder. During training, both N target frames and (N + 1) conditioning frames are passed through the encoder and represented in the latent domain, where the model learns to reconstruct the target latent from the conditioning latent. During inference, only the (N + 1) conditioning frames were encoded, and the model’s predicted latent representations of the N intermediate frames were subsequently decoded back into the pixel space using the decoder of the autoencoder.

Notably, while existing latent diffusion-based VFI models such as LDMVFI and LBBDM often incorporate conditioning frames within their tokenizer stages to facilitate interpolation, our autoencoder processes each frame independently. This design choice ensures a strict decoupling of functional roles: the autoencoder is exclusively responsible for spatial compression and reconstruction, while the denoising network focuses solely on the latent-domain denoising process. By isolating these objectives, we minimize the functional overlap between components, allowing for more targeted optimization of the denoising network’s performance, provided the autoencoder maintains high-fidelity reconstruction.

To determine the most suitable pretrained autoencoder, we compared multiple candidates and considered their reconstruction quality as an approximate upper bound of the VFI performance. Even if a VFI model achieves perfect interpolation in the latent space, the final pixel-level quality ultimately depends on the autoencoder’s decoder, which maps the interpolated latent representations back to the image domain. Hence, insufficient reconstruction capability in the autoencoder would inherently limit the achievable VFI quality.

Based on this assumption, we quantitatively evaluated multiple autoencoders using 2048 randomly sampled videos from both the SNUBH training split and CoronaryDominance dataset. The candidates included: (1) VAE from Stable Diffusion(SD)³⁴ 1.5(temporal window of 8), (2) temporal VAE from Latte²², (3) Cosmos-Tokenizer-CV²¹ (Continuous Video; temporal window of 17), (4) Cosmos-Tokenizer-CI²¹ (Continuous Image), (5) VQ-FIGAN from LDMVFI¹⁶, and (6) VQ Model from LBBDM²⁴. For each candidate, the input frames were first passed through the encoder and subsequently reconstructed using the decoder. The reconstruction quality was then assessed using the same metrics employed for the quantitative evaluation of the interpolation performance.

Cosmos-Tokenizer-CI was ultimately selected as the default autoencoder because it consistently provided a robust reconstruction quality across the datasets. Although the VAE (SD 1.5) yielded slightly better LPIPS scores, qualitative inspection revealed no perceptible visual differences from the human observers, even when the numerical gaps were large. However, reconstructions from the VAE (SD 1.5) occasionally exhibited readily recognizable artifacts such as degraded details around alphabetic characters, catheter tips, and medical devices, resulting in frame-to-frame flickering when viewed as videos.

For frame conditioning, we drew inspiration from the parameter-free conditioning strategy of VIDIM¹⁵. Embeddings of the conditioning and target frames were interleaved in sequence and concatenated before being fed into the first Transformer block. To ensure that the denoising network recognizes the conditioning frames as clean references, we assigned them a minimal noise-level time step. Conditioning timesteps were likewise interleaved with those of the target frames and embedded together as inputs to the Transformer block. In addition to frame conditioning, we integrated ECG signal conditioning to provide information on the cardiac rhythm. Embeddings were extracted from the ECG signal corresponding to the sampled interval and provided as additional conditioning inputs to the Transformer block. However, our experiments showed that incorporating ECG signals did not yield a significant improvement in the interpolation performance, as detailed in Table 3. Therefore, the final Angio-FILM model was implemented without utilizing ECG conditioning at inference.

All the conditioning signals were trained with classifier-free guidance (CFG)³⁵. For frame conditioning, masking was applied at the level of individual frames, implemented by replacing selected frames with the maximum noise in a manner similar to VIDIM. For the ECG signals, masking was performed at the batch level by zeroing out the entire signal embedding.

Following DiT³⁶, we note that the inductive bias of U-Net is not necessarily essential for the performance of the diffusion models. Instead, adopting a standard Transformer-based design allows us to leverage advantages, such as scalability and improved cross-domain adaptation. For the denoising network, we employed variant 1 of Latte²², a video Transformer architecture in which spatial and temporal Transformer blocks are interleaved. The model was randomly initialized and trained from scratch on the dataset.

The training objective was parameterized using a velocity vector field. Although alternative parameterizations are possible—for example, predicting x₁ (clean data) or x₀ (noise) directly—these approaches can suffer from numerical instabilities as t → 1 or t → 0, respectively³⁷. In contrast, simply using the velocity vector field as the training target avoids such issues and has also been adopted in recent large-scale video generation works, such as Movie Gen³⁸. Because our aim is not to reconstruct the latent features of the conditioning frames, the loss is computed only on the latent features of the target frames in a similar manner to MAE³⁹ and TAPNext⁴⁰.

Experimental setup

All the experiments were conducted using four NVIDIA H100 GPUs with a batch size of four per GPU. The training was performed for 500,000 iterations. We adopted a linear learning rate scheduler with a base learning rate of 1e-04 preceded by 10,000 warm-up steps. For the CFG, we applied random masking during training, setting the masking probability for the conditioning frames to 0.1 and the masking probability for the ECG signals to 0.1.

For baseline comparisons, we considered four prior works with different approaches. An optical flow-based method, RIFE, was trained on the same dataset following the original training procedure provided in a publicly available implementation for the Vimeo Triplet dataset. The input size was modified to 512 × 512 to accommodate coronary angiography data. The training was continued until the loss converged. During inference, the model generates one intermediate frame, given two consecutive input images. We also included two generative baselines with similar formulations of the generative method: LDMVFI and LBBDM. Both methods were trained on the same dataset as in our approach to ensure a fair evaluation. As LDMVFI and LBBDM take image triplets as inputs, we randomly sampled three consecutive frames from each video during training. We adopted the autoencoder checkpoints provided by the respective authors and froze them during training rather than during retraining. For the training procedure, we followed the default configurations in the released LDMVFI and LBBDM codes, except for resizing the input images to 512 × 512. Both baselines were trained until convergence, and quantitative evaluations were performed using checkpoints saved after 70 k iterations.

Deterministic sampling was performed by integrating the learned flow ODE using the torchdiffeq library. Consistent with prior observations^19,41 that flow-matching models achieve high quality with relatively few function evaluations, we adopted 50 integration steps for our model in all experiments. For the baseline models, we strictly followed the sampling configurations reported in the original papers (e.g., LDMVFI with DDIM sampling using 200 steps and LBBDM with 50-step sampling).

Evaluation method

The evaluation dataset consisted of two sources: (1) 2048 videos randomly sampled from the SNUBH test split and another (2) 2048 videos from the public CoronaryDominance dataset. The sample size of 2048 videos per dataset was determined in consideration of the quantitative evaluation (see “Quantitative evaluation protocol”), particularly the Fréchet Inception Distance (FID), one of the quantitative metrics used in this study, which is known to be sensitive to sample size and requires sufficiently large sets for reliable estimation⁴². Importantly, our evaluation size is larger than the benchmark test sets commonly used in prior works on natural images^16,24, thereby providing sufficient data for a stable comparison in our medical imaging setting. All subsequent quantitative and human evaluations were conducted using this dataset. The quantitative evaluation used the full set, whereas the human evaluations sampled subsets from it.

Because accurate frame reconstruction is critical in medical VFI, we adopted three widely used full-reference metrics to evaluate the similarity between the interpolated frames and their ground-truth counterparts: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Learned Perceptual Image Patch Similarity⁴³. PSNR provides a per-pixel fidelity measure (higher is better). SSIM evaluates luminance, contrast, and structural similarity using a Gaussian-windowed formulation (higher is better). LPIPS measures the perceptual similarity at the image level by comparing deep features extracted from a pretrained network (lower is better).

To assess the distributional realism, we also report FID computed on frame-level features. FID extracts high-level features using a fixed backbone and compares the reference and generated sets using Fréchet distance. While Inception-v3 trained on ImageNet is commonly used as the fixed backbone, standard CNN architectures tend to overweight textures⁴⁴, which are less informative in coronary angiography. Prior work indicates that this choice of backbone can introduce misalignment with human judgments, and that self-supervised features can mitigate this effect⁴⁵. Accordingly, we chose DINOv2²⁶ as the backbone for computing FID. For a fair comparison, all baseline models and our proposed Angio-FILM were trained and evaluated on the same datasets using the pyiqa⁴⁶ package, which provides standardized implementations of the adopted metrics.

We conducted an expert study to subjectively compare the perceptual quality and angiographic realism of videos generated by Angio-FILM with those produced by baseline VFI models. Four interventional cardiologists, who are also authors of this study (D.-H.K., I.T.M., K.-H.J., and S.-H. Kang) performed a two-stage evaluation to compare perceptual preferences across models using standardized evaluation guidelines. All participants evaluated the same set of data across both stages, ensuring that each item was reviewed by all four experts. Before the evaluation, they shared representative failure cases and discussed visual indicators that were useful for identifying interpolation artifacts. Evaluations were performed without time constraints, and the participants could freely play, pause, and inspect videos at the frame level.

In the first stage, 50 cases were randomly sampled from the SNUBH evaluation dataset. Each case consisted of five videos displayed simultaneously: one original video (explicitly labeled as “original”) and four interpolated videos corresponding to that original video, each generated by a different model and arranged in random order. The raters were asked to select the interpolated video that most closely resembled the original. The selection frequencies were aggregated to characterize the overall preference trends. No statistical inferences were made at this stage; only the frequency of preferences was reported.

In the second stage, 15 videos per method (original, RIFE, LDMVFI, LBBDM, and Angio-FILM) were independently sampled from each of the two evaluation datasets (SNUBH and CoronaryDominance), yielding 150 videos in total. Because sampling was conducted independently for each method, the videos did not form original-interpolated pairs across methods. The participants viewed each video individually and rated its perceptual quality on a 10-point scale (1 = poor, 10 = excellent) based on angiographic reality. Detailed guidelines for quality scoring are provided in Supplementary Table 5. The inter-observer agreement among the four raters was first evaluated using the ICC, where a value greater than 0.75 is generally interpreted as indicative of good agreement. Subsequently, to compare the perceptual quality scores of RIFE, LDMVFI, LBBDM, and Angio-FILM relative to those of the original videos, a linear mixed model (LMM) was applied separately to the SNUBH and CoronaryDominance datasets. The LMM allows for the simultaneous consideration of fixed and random effects, enabling the estimation of mean score differences between models while accounting for rater-level variability. In this analysis, the video generation method was specified as a fixed effect, and the rater was modeled as a random effect. The score of the original videos was defined as the reference level, and the LMM-adjusted marginal mean scores of the four models (RIFE, LDMVFI, LBBDM, and Angio-FILM) were compared against it, accounting for inter-rater variability. A smaller mean difference from the original videos indicates that the corresponding model was perceived as more visually similar to the original, thereby reflecting higher perceptual quality and realism. For each dataset, mean score differences and 95% confidence intervals (95% CI) were computed, and a significance level of 0.05 was used for all statistical tests.

Visual Turing test protocol

The original videos were compared with those interpolated by Angio-FILM. From the SNUBH evaluation pool of 4096 clips (2048 real and 2048 interpolated), we randomly sampled 1200 clips (600 real and 600 interpolated) for human evaluation. Thirty participants were enrolled with stratification by clinical experience: 5 residents, 5 fellows-in-training, 5 non-interventional cardiologists, and 15 interventional cardiologists, with five from each of the following experience levels: 1–5 years, 6–10 years, and ≥10 years. Each participant evaluated a unique and nonoverlapping set of video clips.

The study followed a two-stage protocol. Across both stages, evaluations were performed under uniform conditions: there was no time limit; once a judgment was submitted, the item could not be revisited; participants could play/pause and scrub at the frame level; and each clip was assigned to only one evaluator, ensuring that there was no overlap of clips across participants. All evaluations were conducted using a custom web-based platform developed for this study, which ensured standardized playback control, randomization, and secure response logging. The user interface of this evaluation server is illustrated in Fig. 6, and readers can access the website (https://github.com/SNUBH-CVC/AngioFILM).

In the first stage (single-video binary classification), each participant viewed 20 independent single videos (10 real, 10 interpolated, randomized order) presented one-by-one and judged for each whether it was real or interpolated. This setup closely aligns with the HYPE∞ protocol⁴⁷, such as unlimited viewing time, a 1:1 ratio of real to interpolated videos, and single-stimulus presentation. If a clip was judged to be AI-interpolated, participants additionally rated its perceptual quality on a 10-point scale (1 = poor, 10 = excellent). Accuracy, sensitivity, and specificity were calculated using a generalized linear mixed model (GLMM) to assess the overall performance and differences according to clinical experience. In addition, for videos that participants identified as AI-generated, the perceptual quality scores of the Angio-FILM–interpolated and original videos were compared based on the mean quality scores adjusted for inter-rater variability using LMM.

In the second stage (two-alternative forced choice), each participant viewed ten pairs of videos displayed side by side; each pair contained one real and one interpolated video with randomized left–right positions. Participants were asked to decide which of the two videos was real following a two-alternative forced choice design. The real and interpolated videos forming a pair were not counterparts, meaning that the interpolated video’s original ground truth was not necessarily the real video shown alongside it. Accuracy was analyzed using GLMM, and the model-based probability of a correct response was tested to determine whether it differed significantly from the chance level of 50%. Both the GLMM and LMM frameworks appropriately account for non-independence due to repeated measures and inter-rater variability. Wald statistics were used to evaluate the significance of the regression coefficients estimated from the mixed models, thereby determining whether the model-based probabilities or mean differences differed significantly from the null level, corresponding to random chance. All tests were two-sided, and the significance level was set at 0.05.

QCA standardization and measurement protocol

From the test dataset, 500 video sequences were randomly sampled after excluding cases with an absence of significant stenosis (defined as percent diameter stenosis [%DS] >30%) or unclear contrast agent injection to ensure high measurement reliability. The QCA procedure was strictly standardized: first, a representative frame providing the clearest visualization of the target vessel was selected exclusively from the original video sequences. To facilitate a direct comparison with the model’s output, this selection was restricted to even-numbered frames, which served as the ground truth for the corresponding interpolated frames. QCA was performed using MPXA-2000 software (Medipixel, Seoul, Korea) by an experienced analyst with more than 10 years of expertise. To eliminate potential observer bias, a total of 1000 video frames (500 original and 500 interpolated) were presented in a fully randomized sequence to the analyst, who remained blinded to the frame type. For each analyzed frame, the most severe lesion was identified, and the minimal lumen was measured. Finally, a comparative analysis was performed between the MLD values derived from the original frames and those from the interpolated frames to assess the model’s ability to preserve critical diagnostic geometry.

Author contributions

H.K., D.-H.K. and S.-H.Kang conceived the study. H.K., S.P. and S.-K.K. developed the model and performed experiments. H.K. and S.P. drafted the initial manuscript. S.-H.Kim performed data curation and statistical analysis. I.T.M., K.-H.J. and S.-H.Kang performed human evaluation. C.-H.Y, T.-J.Y. and I.-H.C participated in critical reviewing. S.-H.Kang is responsible for the final manuscript. All authors have read and approved the final manuscript.

Data availability

The datasets generated and/or analyzed during the current study are not publicly available due to patient privacy and institutional data protection policies but are available from the corresponding author on reasonable request. CoronaryDominance data are available at the following URL: https://www.nature.com/articles/s41597-025-04676-8.

Code availability

The custom code and scripts used for the development of Angio-FILM and the analysis of datasets in this study are available on GitHub at https://github.com/SNUBH-CVC/AngioFILM. The model was implemented using Python and PyTorch (specifically utilizing the torchdiffeq library for ODE integration), with specific variables and hyperparameters as detailed in the “Methods” section.

Competing interests

S.-H.K., H.K. and S.P. are inventors on a pending patent application (No. 10-2026-0008121) titled “Apparatus and Method for Interpolating Vascular Angiography Image Frames,” filed by Seoul National University Hospital on January 15, 2026. This patent application covers the latent flow matching-based video frame interpolation methodology and the specific Angio-FILM apparatus described throughout this manuscript. The remaining authors (D.-H.K., S.-K.K., S.-H.Kim, I.T.M., K.-H.J., C.-H.Y., T.-J.Y. and I.-H.C.) declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41746-026-02923-9.