{"id":"44212b29-c6fb-476d-8129-f1f2e68148c2","arxiv_id":"2501.08780","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A residual CNN, adapted from a spatial super-resolution network, temporally upsamples 4D Flow MRI by 2x with lower error than linear or sinc interpolation and generalizes to in-vivo data.","lead":"A neural network can double the number of time frames in 4D Flow MRI heart scans after acquisition, recovering fast blood-flow changes that standard interpolation misses. The method works across synthetic and real patient data without retraining per patient, potentially improving cardiac flow quantification without longer scan times.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"In-vivo evaluations lack linear/sinc baselines; observed correlations could reflect denoising plus interpolation rather than temporal super-resolution.","rationale":"The paper is a competently executed method study: the synthetic pipeline is detailed, the in-silico validation is clean and includes linear/sinc comparators, and the code is public. Those are real strengths. My stress-test focuses on the part of the central claim that goes beyond simulation: the claimed generalization to unseen in-vivo data. The reader's weakest assumption was that the synthetic pipeline faithfully mimics real 4D Flow MRI degradation; I agree that is a risk, but I think the more directly actionable gap is that the in-vivo experiments do not include the same deterministic baselines used in-silico. The recovery-to-native experiment (3.2.1) is particularly weak because deleting frames from already reconstructed data does not reproduce acquisition-time temporal averaging; it tests frame interpolation on a reconstructed sequence. The paired experiment (3.2.2) uses a true low-res acquisition, but the reference is a noisy high-res scan rather than ground truth. Without linear/sinc comparisons on these same inputs, high correlation (k around 0.9) is expected even for trivial interpolation because cardiac flow is temporally smooth and the dynamic range is large. The proposed test is a simple control that the authors can run with their existing code. If interpolation performs as well, the paper should be revised to reposition the contribution as denoising-plus-interpolation; if it performs much worse, the temporal-super-resolution claim is substantially strengthened. This does not change the reader's CONDITIONAL verdict; it sharpens the condition.","tokens_in":14567,"tokens_out":15096,"duration_ms":161613,"concrete_test":"Compute the Table 3/4 metrics (RMSE, k, R2) for linear and sinc interpolation applied directly to the in-vivo low-resolution inputs of both Dataset I (66.8 ms inputs) and Dataset II (40 ms inputs), using the identical fluid masks and peak-frame definitions. If linear interpolation achieves k >= 0.85 and R2 >= 0.90 in the aorta or recovery-to-native peak frames, the network's in-vivo temporal-super-resolution advantage over classical baselines is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Even granting the realism of the Section 2.2.2 synthetic pipeline, the in-vivo evaluations (Sections 3.2.1-3.2.2) contain no deterministic baseline. The recovery-to-native test creates its low-resolution input by removing every second frame from already reconstructed data, which does not reproduce the acquisition-time temporal footprint that the synthetic pipeline is designed to emulate; the paired test compares against a high-resolution acquisition that is noisy and not a gold standard. In neither case are linear or sinc interpolation, the paper's own in-silico comparators, applied to the same in-vivo inputs. Thus the high peak-flow correlations (k=0.90-0.93, R2=0.89-0.95) could be produced by a network that simply denoises and linearly interpolates, without recovering temporal information beyond interpolation. If that were the case, the central claim of clinical temporal super-resolution would be unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a residual convolutional network for 2x temporal super-resolution of 4D Flow MRI, adapting the spatial network 4DFlowNet to operate on 2D+t patches. Training uses paired synthetic data generated from six patient-specific CFD models of the left heart, processed through a pipeline that simulates coil sensitivities, k-space noise, phyllotaxis undersampling, temporal subsampling, and compressed-sensing reconstruction. The loss combines a fluid/non-fluid MSE term with a mutually projected directional L1 loss. Evaluation is performed on one held-out in-silico model and on two in-vivo cohorts (each n=5): one in which low-resolution input is created by removing every second frame from native-resolution data, and one with paired acquisitions at 20 ms and 40 ms temporal resolution. The authors report in-silico RMSE of 1.4 cm/s versus 3.0 cm/s (linear) and 3.8 cm/s (sinc), and in-vivo peak-flow correlations of k=0.93 with R^2=0.89 for recovery-to-native and k=0.90 with R^2=0.95 for paired aortic data.","tokens_in":14697,"tokens_out":4175,"duration_ms":45386,"significance":"If the claims are fully supported, this would be a useful first demonstration of a CNN for temporal super-resolution of 4D Flow MRI that generalizes without patient-specific retraining. The strengths include public release of code and training details, a comparatively realistic synthetic-data pipeline (Section 2.2.2), and an in-silico benchmark against deterministic linear and sinc interpolation that shows a substantial improvement in RMSE and correlation. The main unresolved issue is the strength of the in-vivo evidence: because the in-vivo evaluations omit the same deterministic baselines used in-silico, and because the synthetic-to-real degradation gap is not directly quantified, the paper's central clinical claim is not yet established at the level the abstract implies.","major_comments":[{"comment":"The in-vivo evaluations contain no linear or sinc interpolation baselines applied to the same low-resolution inputs. In the absence of such baselines, the reported peak-flow correlations (k=0.90–0.93, R^2=0.89–0.95) could be produced by a network that mainly denoises and linearly interpolates, without recovering temporal information beyond interpolation. Please add deterministic interpolation results for the recovery-to-native and paired in-vivo datasets, and report the same metrics (RMSE, k, R^2) for those baselines.","section":"§3.2, Tables 3–4"},{"comment":"The recovery-to-native evaluation constructs low-resolution input by removing every second frame from already reconstructed native-resolution data. This does not reproduce the temporal footprint that the synthetic pipeline in §2.2.2 is designed to emulate, where temporal subsampling occurs through k-space readout accumulation followed by compressed-sensing reconstruction. The test therefore measures frame interpolation of reconstructed frames rather than recovery from an acquisition-like temporal degradation. Please either create low-resolution in-vivo inputs with a comparable temporal footprint (e.g., retrospective temporal binning or re-sampling of k-t data) or explicitly quantify and discuss the effect of this mismatch on the reported generalization.","section":"§2.2.4 and §3.2.1"},{"comment":"The in-silico test set consists of a single held-out CFD model, even though it is divided into many patches. The claim of generalization to an 'unseen in-silico setting' is therefore based on n=1 at the model level. Please report results for multiple held-out models or provide model-level confidence intervals, and discuss how patch-level statistics may overstate the diversity of the test set.","section":"§2.2.3 and §3.1"},{"comment":"The paired in-vivo evaluation compares super-resolved data against a separately acquired high-resolution dataset that is itself noisy and may differ in heart rate, breathing, and subject position. The manuscript acknowledges this in §4.5, but the quantitative metrics are presented as if the high-resolution reference were a gold standard. Please report the estimated noise level of the high-resolution reference (e.g., SNR or test–retest variability) and discuss how reference noise bounds the achievable k and R^2 values.","section":"§3.2.2 and §4.5"}],"minor_comments":[{"comment":"The caption lists the input velocities as '˜vx, ˜vx, ˜vx' and the outputs as 'ˆvx, ˆvx, ˆvx'; these should be vx, vy, vz and vx, vy, vz respectively.","section":"Figure 1 caption"},{"comment":"The relationship between the relative error RE defined in Eq. (1) and the MRE values reported in Table 2 is not specified. Please state explicitly whether MRE is the average of RE over voxels and time, and how the percentage is obtained.","section":"§2.3.3, Eq. (1), Table 2"},{"comment":"RMSE values are reported as triples for the three velocity components while MAE and MRE are single numbers. Please specify whether MAE is computed on velocity magnitude or averaged over components, so the table is self-consistent.","section":"Table 2"},{"comment":"The preliminary comparison between the proposed pipeline and a simpler 'remove every second frame plus k-space noise' pipeline is described qualitatively as improving in-vivo performance. Please provide quantitative results for this comparison or clearly mark it as anecdotal.","section":"§4.2"},{"comment":"The text states that target SNR is 'randomly assigned between 14-17 dB or 40-45 db (two separate datasets)', but the validation and test sets are described as high-noise only. Please clarify which SNR ranges are used for training versus validation/test and whether the test model was exclusively high-noise.","section":"§2.2.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of a medical imaging or machine-learning journal, and the open-source code is a positive feature. The main risk is not circularity—the synthetic training and held-out evaluation design is sound—but that the in-vivo evaluation design does not yet rule out the trivial explanation of denoising plus interpolation. Adding in-vivo linear/sinc baselines and addressing the degradation mismatch in the recovery-to-native experiment would substantially strengthen the paper. The in-silico n=1 model-level test is another concern that should be addressed before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline: this is a credible proof-of-principle for temporal super-resolution in 4D Flow MRI. The authors adapt 4DFlowNet from spatial to temporal upsampling, build a synthetic data pipeline that mimics k-space sampling and compressed sensing, and test on a held-out CFD model plus two small in-vivo cohorts. The in-silico result is strong: RMSE 1.4 cm/s vs 3.0-3.8 for linear/sinc interpolation. That's a real improvement.\n\nWhat's genuinely new: the temporal adaptation itself, and the pipeline for generating realistic low-resolution training data. Most prior DL work in 4D Flow has been spatial SR, and the authors are honest that their architecture is borrowed and that LSTM didn't help. They also release code, which helps the field.\n\nSoft spots: the in-vivo evaluation is the main weakness. In the paired dataset, the SR output is compared only against a noisy high-res acquisition, with no linear or sinc interpolation baselines. So the impressive correlations (k=0.90-0.93) could in principle come from a network that denoises and linearly interpolates rather than recovering true temporal information. The recovery-to-native test creates low-res input by removing every second frame from reconstructed data, which doesn't reproduce the acquisition-time temporal footprint that the synthetic pipeline tries to emulate. The stress-test's concern is valid on this point.\n\nThat said, the in-silico comparison to interpolation already shows the network does more than interpolation on synthetic data, and the synthetic pipeline is a reasonable attempt to bridge the gap. The real issue is that the in-vivo evidence per se doesn't support the clinical claim as strongly as the abstract suggests. The paper's discussion is more measured, but the abstract overstates it.\n\nOther minor issues: only one held-out CFD model for the main test, no error bars or significance testing, and the code lacks training data for full reproducibility. The authors acknowledge the small dataset and the lack of learned baselines in the limitations.\n\nBottom line: this is a useful first step and a reasonable baseline for others. It deserves peer review and I'd expect heavy but surmountable revision. The authors should add in-vivo interpolation baselines, report variance across subjects, and temper the abstract's 'enabling' language to 'potentially enabling'. If they do that, it's a solid methods paper.","headline":"A solid proof-of-principle for temporal super-resolution in 4D Flow MRI that overclaims in the abstract but is worth engaging with after adding in-vivo baselines.","tokens_in":15297,"tokens_out":2823,"would_cite":true,"duration_ms":27281,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A residual CNN temporally super-resolves 4D Flow MRI, recovering flow peaks that linear and sinc interpolation miss.","keywords":["temporal super-resolution","4D Flow MRI","residual neural network","deep learning","compressed sensing","computational fluid dynamics","cardiac flow quantification","denoising"],"falsifier":"A concrete disconfirming experiment would be to acquire paired 20 ms and 40 ms 4D Flow MRI in patients with cardiac pathology outside the mitral regurgitation cohort and compute the network's RMSE on frames during late diastole or atrial contraction; if the error approaches or exceeds the linear-interpolation error on those frames, the claim of generalizable temporal super-resolution is refuted.","tokens_in":14321,"feed_emoji":"🫀","tokens_out":7585,"duration_ms":68005,"temperature":0.7,"pith_summary":"This paper asks whether a convolutional network trained on synthetic data can increase the temporal resolution of 4D Flow MRI, a technique whose clinical use is limited by the trade-off between acquisition time, noise, and frame rate. The authors adapt a spatial super-resolution network, 4DFlowNet, to upsample in time using 2D+t patches, and train it on low-resolution inputs generated from patient-specific computational fluid dynamics via a pipeline that mimics real MRI k-space sampling, noise, and compressed-sensing reconstruction. On a held-out in-silico model the network reaches a mean absolute error of 1.0 cm/s versus 2.3–2.6 cm/s for linear and sinc interpolation, and on two in-vivo cohorts it synthesizes missing frames with linear-regression slopes around 0.9 at peak systole. The point of the result is that high-frame-rate volumetric flow quantification could be obtained from acquisitions of roughly half the length, without patient-specific retraining.","feed_headline":"CNN doubles 4D Flow MRI frame rate, beating interpolation on peak flow","feed_subtitle":"Trained on simulated MRI, the network recovers cardiac frames the scanner never captured, promising shorter scans.","key_machinery":"The load-bearing mechanism is a residual convolutional network derived from 4DFlowNet, in which the central upsampling layer was changed from trilinear to linear to increase the frame count only along the temporal axis; inputs are 16x16x16 patches of 2D spatial slices sampled over 16 time frames, and outputs are 16x16x32 patches at double the frame rate. The loss combines an MSE term computed separately over fluid and non-fluid voxels with the mutually projected loss, an angle- and magnitude-aware L1 measure from SRflow, plus an L2 weight regularizer. The second half of the machinery is the data-generation pipeline: CFD velocities are converted to phase-MRI signals, multiplied by simulated Biot-Savart coil sensitivities, corrupted with k-space noise, subsampled with a variable-density phyllotaxis pattern that accumulates consecutive frames to emulate temporal undersampling, and reconstructed by compressed sensing with FISTA in BART. That pipeline is what gives the network inputs whose temporal-smoothing and noise artifacts resemble clinically acquired 4D Flow MRI.","core_discovery":"The paper's central claim is that temporal super-resolution of 4D Flow MRI is tractable with a residual CNN: given three-dimensional spatial slices sampled over time at low frame rate, the network outputs denoised velocities at double the frame rate, synthesizing information that is absent in the low-resolution input. In the in-silico test the super-resolved velocities match the high-resolution reference with an average RMSE of 1.4 cm/s inside the fluid region (and 0.2 cm/s outside), against 3.0 cm/s for linear and 3.8 cm/s for sinc interpolation. In-vivo, the network recovers synthetically removed frames in five subjects with k=0.93 and $R^{2}$=0.89 at peak systole, and in a paired low/high-rate acquisition in five further subjects it super-resolves 40 ms input to 20 ms reference data with k=0.90 and $R^{2}$=0.95 in the aorta. The authors frame the work as a first-of-a-kind temporal super-resolution baseline for 4D Flow MRI that generalizes across anatomies without retraining.","pith_inferences":["We infer that the temporal smoothing in current 4D Flow MRI is partly an artifact of undersampling and reconstruction rather than truly absent information, since a network trained on realistic synthetic data reconstructs peaks not present in the low-rate input.","A testable extension is to ablate the synthetic pipeline stages (coil sensitivities, phyllotaxis sampling, compressed sensing, temporal accumulation) to determine which component is responsible for the in-vivo generalization; the paper only notes that removing the realistic pipeline degraded performance, not which stage matters most.","The focus on left-ventricular geometries with mitral regurgitation suggests the approach may need additional training data from other cardiovascular compartments before it becomes a general clinical tool, and the method could be extended to those domains using the same pipeline.","If the result transfers, the same 2D+t residual design could be applied to other time-resolved flow or cine imaging modalities, with the main challenge being construction of a similarly realistic synthetic training pipeline."],"forward_implications":["Post-processing could raise the effective temporal resolution of existing 4D Flow MRI datasets from roughly 40 ms toward 20 ms without rescanning, since the network generates intermediate frames.","Peak-flow estimates (early diastole and peak systole) become more accurate: the in-silico RMSE at those frames is 1.5–1.7 cm/s, several-fold lower than for interpolation.","The network simultaneously denoises the input, reducing error in non-fluid regions to 0.2 cm/s, which could improve downstream segmentation and visualization.","The method requires no patient-specific retraining, no predefined fluid-region segmentation, and runs patch-based inference with stitching, so it can be applied to retrospective clinical data.","Generalization across two in-vivo protocols with different spatial resolutions and temporal resolutions indicates the network is not overfit to a single acquisition sequence."],"supporting_citations":[{"why":"Supplies the base convolutional architecture, 4DFlowNet, that the paper re-designs for temporal upsampling.","marker":"[18]"},{"why":"Earlier spatial super-resolution work by the same group whose network and fluid-boundary error behavior inform the design.","marker":"[7]"},{"why":"Provides the patient-specific CFD simulations of mitral regurgitation that yield the high-resolution velocity ground truth.","marker":"[13]"},{"why":"Supplies the variable-density phyllotaxis k-space sampling pattern used to mimic real acquisitions in the synthetic pipeline.","marker":"[23]"},{"why":"Establishes the compressed-sensing formulation used in the reconstruction step of the synthetic pipeline.","marker":"[24]"},{"why":"Provides the FISTA optimization algorithm used to solve the compressed-sensing reconstruction.","marker":"[25]"},{"why":"Supplies the BART implementation used for the compressed-sensing reconstruction.","marker":"[26]"},{"why":"Contributes the mutually projected loss function incorporated into the training loss.","marker":"[19]"}],"fun_headline_variants":["Neural net adds missing MRI time frames, beating linear interpolation","CNN super-resolves 4D Flow MRI in time, not just space","Deep learning doubles cardiac MRI frame rate from low-res scans","AI synthesizes absent 4D Flow MRI frames, outperforming interpolation","Temporal super-resolution: CNN upsamples 4D Flow MRI beyond interpolation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the synthetic pipeline produces low-resolution 4D Flow MRI inputs whose noise and undersampling artifacts match real clinical acquisitions, so that in-vivo performance reflects true recovery rather than the network learning to denoise generic flow patterns.","fun_headline_variants_meta":{"raw":{"variants":["Neural net adds missing MRI time frames, beating linear interpolation","CNN super-resolves 4D Flow MRI in time, not just space","Deep learning doubles cardiac MRI frame rate from low-res scans","AI synthesizes absent 4D Flow MRI frames, outperforming interpolation","Temporal super-resolution: CNN upsamples 4D Flow MRI beyond interpolation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000726,"raw_usage":{"total_tokens":3328,"prompt_tokens":1094,"completion_tokens":2234,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":710,"completion_tokens_details":{"reasoning_tokens":2141}},"tokens_in":710,"tokens_out":2234,"duration_ms":15249,"temperature":1.0,"reasoning_tokens":2141,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:17:22.421427+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete disconfirming experiment would be to acquire paired 20 ms and 40 ms 4D Flow MRI in patients with cardiac pathology outside the mitral regurgitation cohort and compute the network's RMSE on frames during late diastole or atrial contraction; if the error approaches or exceeds the linear-interpolation error on those frames, the claim of generalizable temporal super-resolution is refuted.","supporting_citations":[{"cited_title":"Dubowitz, Debbie Zhao, Alan Wang, Brett Cowan, and Alistair A","cited_arxiv_id":null,"evidence_quote":"Supplies the base convolutional architecture, 4DFlowNet, that the paper re-designs for temporal upsampling."},{"cited_title":"Cerebrovascu- lar super-resolution 4D Flow MRI-using deep learning to non-invasively quantify velocity, flow, and relative pressure Cerebrovascular super-resolution 4D Flow MRI","cited_arxiv_id":null,"evidence_quote":"Earlier spatial super-resolution work by the same group whose network and fluid-boundary error behavior inform the design."},{"cited_title":"Bonini, M","cited_arxiv_id":null,"evidence_quote":"Provides the patient-specific CFD simulations of mitral regurgitation that yield the high-resolution velocity ground truth."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the variable-density phyllotaxis k-space sampling pattern used to mimic real acquisitions in the synthetic pipeline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the compressed-sensing formulation used in the reconstruction step of the synthetic pipeline."},{"cited_title":"A fast iterative shrinkage- thresholding algorithm for linear inverse problems","cited_arxiv_id":null,"evidence_quote":"Provides the FISTA optimization algorithm used to solve the compressed-sensing reconstruction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the BART implementation used for the compressed-sensing reconstruction."},{"cited_title":"Paetzold, Augusto F","cited_arxiv_id":null,"evidence_quote":"Contributes the mutually projected loss function incorporated into the training loss."}],"review_version":1}