{"id":"1431d74f-29d1-4a6e-ac1b-c0fd4571f4f3","arxiv_id":"1908.08947","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A 3D Dense-UNet with multi-level SSIM loss reconstructs 4D perfusion and angiography from 50%-sampled interleaved crushed and non-crushed Hadamard te-pCASL data, achieving SSIM 97.3% (perfusion) and 96.2% (angiography) on synthetic test data.","lead":"An MRI reconstruction network reconstructs both 4D blood-flow (perfusion) and angiography images from half the usual Hadamard-encoded ASL scan data, cutting acquisition time in half. The authors train it on synthetic data generated from a kinetic model; clinical validation on real patients remains future work.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim rests on synthetic training data; without in vivo validation, reported SSIM may not transfer to real te-pCASL acquisitions.","rationale":"The reader's weakest_assumption identifies exactly the synthetic-to-real generalization gap, and the full paper text confirms it: the data generator is the authors' own kinetic model, all test sets are generated from that model, and the Conclusion explicitly defers in vivo data to future work. This is the single most load-bearing concern because the headline contribution is a factor-of-two reduction in clinical scan time; if the network only learns to invert the simulator, the reported SSIM values do not establish that the method works in patients. I considered whether an internal inconsistency exists instead, such as the typographical issues in Eq. (3) or the questionable 'T1b' notation, but these are minor and do not affect the central argument. I also considered the lack of non-CNN acceleration baselines; while relevant for positioning, it is secondary to the absence of real-data validation. The proposed concrete test is a direct external validation: apply the existing trained network to real fully-sampled data subsampled to the same 50% interleaved pattern. This would settle whether the simulator-trained model transfers. The reader's CONDITIONAL verdict remains appropriate, so no verdict change is needed.","tokens_in":8939,"tokens_out":3429,"duration_ms":34865,"concrete_test":"Acquire full 8-row crushed and 8-row non-crushed te-pCASL scans from at least 10 healthy subjects on the same scanner/protocol, compute ground-truth 4D perfusion and angiography by standard Hadamard decoding, subsample the acquired rows to the interleaved 50% pattern used in Eq. (2), and feed these real subsampled volumes through the trained ML-SSIM network. Compare the network output to the full-decoded ground truth using SSIM, SNR, and pSNR. If mean perfusion SSIM drops by more than about 3 points or the angiography pSNR variance grows substantially beyond the reported 35.4 +/- 23.2, the synthetic-only evidence is insufficient to support the claimed clinical acceleration.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that the synthetic te-pCASL data generated from Eqs. (3)-(4) are representative enough of real crushed and non-crushed Hadamard te-pCASL acquisitions for the reported 97.3 +/- 1.1 perfusion SSIM and 96.2 +/- 11.1 angiography SSIM to transfer to clinical use. Every training, validation, and test example comes from the authors' own simulator, which uses only 6 BAT maps, 4 AAT maps, and 20 BrainWeb anatomies. The kinetic model omits transit-time dispersion, partial-volume effects, motion, B1/B0 inhomogeneities, T1/T2 variations, labeling-efficiency errors, and realistic scanner noise; the only noise augmentation is white Gaussian noise added during training. The network can therefore exploit simulator-specific correlations rather than learn physics that generalizes to actual acquisitions. The paper itself acknowledges this limitation in the Conclusion: 'A further step of this study is enriching the training and validation datasets with in vivo data.' The validation step, comparing the generated signals to the Buxton curve model, is also self-referential because the generator is built from the same Buxton model. This absence of real-data validation is the key gap between the in-silico demonstration and the abstract's clinical claim of 'negating the additional scan time.' Disagreement with clinical consensus is not the issue; the correctness risk is that the reported metrics estimate performance on a simulator, not on the target acquisition protocol.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Hadamard time-encoded pCASL can provide both dynamic perfusion and 4D angiography when acquired with and without flow-crushing, but the two acquisitions double scan time. This paper proposes a 3D Dense-UNet that inputs interleaved 50%-sampled crushed and non-crushed Hadamard te-pCASL images (four of eight encodings each) and outputs seven-timepoint perfusion and angiographic volumes, achieving a factor-two acceleration relative to full sampling. Training and test data are generated entirely from a simulator built on a generalized Buxton kinetic model (Eqs. 3-4) using 6 BAT maps, 4 AAT maps, and 20 BrainWeb anatomies. The authors compare MSE, VGG perceptual, SSIM, and their proposed multi-level SSIM losses and report best perfusion SSIM of 97.3 +/- 1.1 and angiography SSIM of 96.2 +/- 11.1 on 313 synthetic test datasets. The paper claims this obviates the additional scan time in practice.","tokens_in":9301,"tokens_out":7590,"duration_ms":76028,"significance":"The contribution is potentially useful: an end-to-end 3D CNN with dense connectivity and a multi-level SSIM loss is a reasonable architecture for recovering both smooth perfusion and sparse vascular structures from undersampled Hadamard-encoded ASL, and the loss-function comparison with Wilcoxon tests is a strength. The data-generation pipeline is clearly specified and makes the method reproducible. The significance is conditional because the entire evaluation is in-distribution with respect to the authors' own simulator: no real crushed/non-crushed te-pCASL data are used for validation, and the metrics therefore estimate performance on synthetic signals that follow Eqs. (3)-(4), not on clinical acquisitions with dispersion, motion, partial volume, or B1/B0 effects. The abstract's clinical claim of 'negating the additional scan time' is stronger than the evidence supports.","major_comments":[{"comment":"All training, validation, and test examples are generated by the authors' simulator from 6 BAT maps, 4 AAT maps, and 20 BrainWeb subjects; the only corruption added is Gaussian white noise with standard deviation between 0 and 5 (Section 3). The kinetic model omits transit-time dispersion, partial-volume effects, motion, B1/B0 inhomogeneities, T1/T2 variation, and labeling-efficiency errors. Because the network is trained and tested on this same distribution, the reported SSIM values (97.3 +/- 1.1 perfusion, 96.2 +/- 11.1 angiography) do not by themselves support the abstract's statement that the method 'negat[es] the additional scan time' in real acquisitions. The paper itself acknowledges this in the Conclusion: 'A further step of this study is enriching the training and validation datasets with in vivo data.' I ask the authors to either add an in vivo or at least a domain-shift experiment (e.g., realistic noise/artifact model or a held-out set from a different simulator/scanner) and to temper the abstract and conclusion so the claim is limited to the synthetic setting.","section":"§2.3, Eqs. (3)-(4); Abstract; Conclusion"},{"comment":"The sentence 'To evaluate the generated data, the signal evolution pattern was validated by the Buxton curve model [15]' is self-referential, because Eqs. (3)-(4) are themselves a generalization of the Buxton model. Matching the Buxton curve therefore only checks internal consistency of the generator, not agreement with real te-pCASL signals. This sentence should be removed or replaced by a comparison with measured crushed/non-crushed te-pCASL data.","section":"§2.3, final sentence"},{"comment":"The paper states the goal of 'accurate CBF quantification' but no quantitative CBF accuracy is reported. Table 1 and Fig. 3 contain only image-space metrics (SSIM, MSE, SNR, pSNR) on reconstructed perfusion volumes; there is no comparison of estimated CBF values (e.g., bias and limits of agreement in mL/100 g/min) against the known simulation parameters or against a reference method. Either add such an analysis or remove the CBF-quantification claim from the introduction and conclusion.","section":"§1 and §4"}],"minor_comments":[{"comment":"For the angiography rows with high variance (SSIM and ML-SSIM), the reported mean +/- SD is strongly skewed (ML-SSIM SSIM mean 96.2, median 99.7, SD 11.1); report median and IQR and state how many of the 313 test datasets drive the low mean.","section":"Table 1, Fig. 3"},{"comment":"The formula 'w1 x Loss level1 + w2 x Loss level1 + w3 x Loss level3' appears to have a typo; the second term should likely be 'w2 x Loss level2'.","section":"Figure 1"},{"comment":"Reference [1] (Ferlay et al., cancer epidemiology) does not correspond to the first sentence about ASL being non-invasive; please replace it with the intended ASL review or methods citation.","section":"Introduction, Reference [1]"},{"comment":"The Gaussian noise augmentation is specified as 'random standard deviation between 0 and 5' without stating the units or the signal scale to which it is applied; please clarify (e.g., as a percentage of the mean signal or in arbitrary intensity units).","section":"§2.3, Section 3"}],"recommendation":"major_revision","confidential_remarks":"This is a competent in-silico proof-of-concept. The main risk is overclaiming a clinical benefit from simulated-data metrics. If the authors add one in vivo or domain-shift validation and calibrate the abstract and conclusion, the paper would be acceptable; otherwise the evidence gap is substantial."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know that this paper is a solid in-silico proof-of-concept, not a clinical result. The headline claim—that interleaved 50% sampling 'negates the additional scan time'—is only demonstrated in simulation, and the abstract overstates it.\n\nWhat is genuinely new: the combination of interleaved half-sampled crushed/non-crushed Hadamard te-pCASL with a single 3D Dense-UNet that outputs both 4D perfusion and 4D angiography in one pass. The extension of the Buxton kinetic model to an explicit arterial compartment (Eq. 3) is a real addition over prior te-ASL models. The internal evaluation is also consistent: ML-SSIM beats MSE, SSIM, and perceptual loss on the test set, and the Wilcoxon tests support that ranking. The data generator is described in enough detail that an interested group could re-implement it.\n\nThe load-bearing weakness is the synthetic data. Every training, validation, and test example comes from the authors' own kinetic model, built from only 6 BAT maps, 4 AAT maps, and 20 BrainWeb anatomies. The validation against the Buxton curve is self-referential because the generator uses that same model. Real acquisitions will include transit-time dispersion, partial-volume effects, motion, B1/B0 inhomogeneity, labeling-efficiency errors, and structured noise—none of which are modeled beyond additive white Gaussian noise. So the reported SSIM of 97.3 ± 1.1 (perfusion) and 96.2 ± 11.1 (angiography) estimates performance on a simulator, not on the target clinical protocol. The authors acknowledge this in the conclusion, but the abstract's clinical framing is not appropriately hedged.\n\nThere are also reproducibility gaps: no code, no data, no loss weights, no kinetic parameter values, and no comparison against a non-CNN acceleration baseline (e.g., compressed sensing or low-rank reconstruction). Minor textual issues exist (Eq. 3 looks like it has a typo; Fig. 1 labels the second ML-SSIM term as 'Loss level1' twice), but these are trivial.\n\nThese flaws are addressable. The central idea is sound and the engineering is clean. Who is this for? Researchers working on ASL acceleration or MRI reconstruction who want to see the first demonstration of joint perfusion/angiography reconstruction from interleaved half-sampled te-ASL. It deserves a serious referee, not a desk reject. I would send it to peer review with the expectation of heavy revision: tone down the abstract, add at least a small in vivo pilot (even a couple of subjects), release code or detailed parameter settings, and include a classical reconstruction comparison. That would turn a promising simulation into a credible method.\n\nFor your own work: I would cite the kinetic model extension and the interleaved acquisition idea, but not the quantitative SSIM claims until they are shown on real data.","headline":"A well-executed in-silico proof-of-concept for 3D CNN reconstruction of perfusion and angiography from half-sampled Hadamard te-pCASL, but the abstract's clinical claim outruns the evidence.","tokens_in":9839,"tokens_out":1880,"would_cite":true,"duration_ms":20493,"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 3D Dense-UNet with multi-level SSIM loss reconstructs 4D perfusion and 4D angiography from interleaved half-sampled crushed and non-crushed Hadamard te-pCASL data, preserving image quality while halving scan time.","keywords":["Hadamard time-encoded ASL","pseudo-continuous arterial spin labeling","4D MRA","4D perfusion","3D convolutional neural network","Dense-UNet","multi-level SSIM loss","MRI acceleration"],"falsifier":"Acquire real crushed and non-crushed Hadamard-8 te-pCASL scans on a cohort, form full-sampled ground truth by the standard decoding and subtraction, then feed the interleaved 50-percent-sampled subset through the trained network; if the SSIM against real full-sampled output falls substantially below 97 percent for perfusion or 96 percent for angiography, the central claim fails. A cheaper check is to compare the kinetic model's predicted arterial and tissue signals to measured signals in voxels with known transit times.","tokens_in":8680,"feed_emoji":"🧠","tokens_out":7499,"duration_ms":64472,"temperature":0.7,"pith_summary":"This paper tries to establish that a single 3D convolutional network can replace a second MRI acquisition. In Hadamard time-encoded pCASL, dynamic perfusion and dynamic angiography are normally acquired separately, one with flow-crushing and one without, which doubles scan time. The authors propose an end-to-end 3D Dense-UNet that takes interleaved half-sampled crushed and non-crushed data and outputs both 4D perfusion and 4D angiography at seven time-points. Training data are generated synthetically from a generalized kinetic model using in vivo arrival-time maps and simulated brain anatomy. On 313 test datasets the network reaches SSIM 97.3 ± 1.1 for perfusion and 96.2 ± 11.1 for angiography, arguing the factor-of-two scan-time penalty is avoidable.","feed_headline":"3D CNN rebuilds 4D perfusion and MRA from half the MRI data","feed_subtitle":"Interleaved half-sampled crushed and non-crushed te-pCASL yields both image types at full quality, cutting scan time.","key_machinery":"The load-bearing device is a 3D Dense-UNet with a multi-level loss. Dense blocks with $3\\times3\\times3$ and $1\\times1\\times1$ convolutions improve feature propagation and gradient flow; up-sampling uses a constant trilinear resize kernel to avoid checkerboard artifacts. The multi-level loss ML-SSIM weights the SSIM loss at three network levels, chosen because SSIM outperformed MSE, VGG-16 perceptual loss, and single-scale SSIM for perfusion reconstruction. The training data come from a generalized Buxton kinetic model (Eqs. 3 and 4) that combines an intravascular arterial signal and an extravascular tissue signal parameterized by AAT, BAT, and CBF; fully sampled Hadamard decoding of the generated scans provides the ground truth.","core_discovery":"The central claim is that a 3D Dense-UNet trained with a multi-level SSIM loss can decode 4D perfusion and 4D angiography directly from interleaved 50-percent-sampled crushed and non-crushed Hadamard te-pCASL scans, without a separate full acquisition. The network uses an 8-row Hadamard encoding, takes four crushed and four non-crushed patches as input, and outputs 14 patches, seven perfusion and seven angiography time-points. With the ML-SSIM loss it achieves SSIM of 97.3 ± 1.1 on perfusion and 96.2 ± 11.1 on angiography for 313 test datasets, and reconstructs all time-points in about 205 ms. The paper also presents a data-generation framework based on the generalized Buxton kinetic model to create realistic training data from in vivo arterial arrival time, bolus arrival time, and simulated CBF maps.","pith_inferences":["The interleaved half-sampling idea could transfer to other dual-acquisition protocols where two scans are subtracted, with the network learning the unmixing directly from data.","Because angiography is intrinsically sparse and dominated by elongated structures, explicit sparsity-aware or structure-aware loss terms could push early time-point angiography quality beyond the reported SNR.","The synthetic-to-real gap is the largest open risk; fine-tuning on a small set of real in vivo crushed and non-crushed pairs would test whether the reported SSIM translates to clinical acquisitions.","With inference under a quarter of a second, the network could be embedded into a real-time reconstruction loop, displaying both perfusion and angiography immediately after acquisition."],"forward_implications":["If the central claim holds, simultaneous 4D MRA and perfusion can be acquired in one interleaved scan rather than two, halving the te-pCASL scan time.","The network reconstructs all seven time-points of both output types in roughly 205 ms per patch set, suggesting a speed compatible with clinical workflow.","The ML-SSIM loss gives a statistically significant improvement over MSE loss for perfusion ($p<0.05$), while matching the SSIM loss for angiography; this identifies the loss design as a key factor.","The synthetic data-generation framework extends to higher-rank Hadamard matrices and other time-encoded ASL schemes, since the kinetic model equations generalize beyond Hadamard-8."],"supporting_citations":[{"why":"Provides the Hadamard te-pCASL decoding scheme used to obtain dynamic perfusion ground truth from crushed scans.","marker":"[2]"},{"why":"Establishes how non-crushed decoded data and perfusion subtraction yield 4D MRA and arterial input function information.","marker":"[3]"},{"why":"Supplies the general kinetic model that the paper generalizes into Eqs. 3 and 4 for generating synthetic te-pCASL signals.","marker":"[15]"},{"why":"Supplies simulated brain anatomy and tissue segmentation used to assign CBF values for training data.","marker":"[16]"},{"why":"Provides the registration method that aligns in vivo arrival-time maps to the simulated anatomy.","marker":"[17]"},{"why":"Gives the Hadamard decoding procedure used to create ground-truth perfusion and angiography from generated scans.","marker":"[18]"},{"why":"Provides the dense connectivity pattern that improves gradient flow and feature reuse in the network.","marker":"[10]"},{"why":"Provides the U-shaped fully convolutional DenseNet architecture that the network is based on.","marker":"[11]"},{"why":"Supplies the trilinear resize up-sampling strategy that prevents checkerboard artifacts.","marker":"[13]"}],"fun_headline_variants":["Half the MRI data, full 4D perfusion and MRA via 3D CNN","3D CNN turns half-sampled ASL into full 4D perfusion and angiography","One scan yields both 4D perfusion and MRA using deep learning","Deep learning cuts ASL scan time in half for dual perfusion and MRA","3D Dense-UNet decodes 4D perfusion and angiography from 50% data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The synthetic te-pCASL data produced by the kinetic model from in vivo arrival-time maps and simulated anatomies is representative enough of real crushed and non-crushed Hadamard te-pCASL acquisitions; if real signals contain effects the model omits, the trained network's outputs may not match the reported SSIM.","fun_headline_variants_meta":{"raw":{"variants":["Half the MRI data, full 4D perfusion and MRA via 3D CNN","3D CNN turns half-sampled ASL into full 4D perfusion and angiography","One scan yields both 4D perfusion and MRA using deep learning","Deep learning cuts ASL scan time in half for dual perfusion and MRA","3D Dense-UNet decodes 4D perfusion and angiography from 50% data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000735,"raw_usage":{"total_tokens":3314,"prompt_tokens":1004,"completion_tokens":2310,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":2213}},"tokens_in":620,"tokens_out":2310,"duration_ms":14112,"temperature":1.0,"reasoning_tokens":2213,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:20:01.136647+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Acquire real crushed and non-crushed Hadamard-8 te-pCASL scans on a cohort, form full-sampled ground truth by the standard decoding and subtraction, then feed the interleaved 50-percent-sampled subset through the trained network; if the SSIM against real full-sampled output falls substantially below 97 percent for perfusion or 96 percent for angiography, the central claim fails. A cheaper check is to compare the kinetic model's predicted arterial and tissue signals to measured signals in voxels with known transit times.","supporting_citations":[{"cited_title":"Advances in arterial spin labelling mri meth- ods for measuring perfusion and collateral flow.Journal of Cerebral Blood Flow & Metabolism, 38(9):1461–1480, 2018","cited_arxiv_id":null,"evidence_quote":"Provides the Hadamard te-pCASL decoding scheme used to obtain dynamic perfusion ground truth from crushed scans."},{"cited_title":"The quasar repro- ducibility study, part II: Results from a multi-center arterial spin labeling test– retest study","cited_arxiv_id":null,"evidence_quote":"Establishes how non-crushed decoded data and perfusion subtraction yield 4D MRA and arterial input function information."},{"cited_title":"A general kinetic model for quantitative perfusion imaging with arterial spin labeling.MRM, 40(3):383–396, 1998","cited_arxiv_id":null,"evidence_quote":"Supplies the general kinetic model that the paper generalizes into Eqs. 3 and 4 for generating synthetic te-pCASL signals."},{"cited_title":"Brainweb: Online interface to a 3d mri simulated brain database","cited_arxiv_id":null,"evidence_quote":"Supplies simulated brain anatomy and tissue segmentation used to assign CBF values for training data."},{"cited_title":"Elastix: a toolbox for intensity-based medical image registration","cited_arxiv_id":null,"evidence_quote":"Provides the registration method that aligns in vivo arrival-time maps to the simulated anatomy."},{"cited_title":"Transit time mapping in the mouse brain using time- encoded pcasl","cited_arxiv_id":null,"evidence_quote":"Gives the Hadamard decoding procedure used to create ground-truth perfusion and angiography from generated scans."},{"cited_title":"Densely connected convolutional networks","cited_arxiv_id":null,"evidence_quote":"Provides the dense connectivity pattern that improves gradient flow and feature reuse in the network."},{"cited_title":"The one hundred layers tiramisu: Fully convolutional densenets for semantic segmentation","cited_arxiv_id":null,"evidence_quote":"Provides the U-shaped fully convolutional DenseNet architecture that the network is based on."},{"cited_title":"Image super- resolution using deep convolutional networks.TPAMI, 38(2):295–307, 2015","cited_arxiv_id":null,"evidence_quote":"Supplies the trilinear resize up-sampling strategy that prevents checkerboard artifacts."}],"review_version":1}