{"id":"6ce34552-4021-45bf-97ec-6e63094ee62d","arxiv_id":"2411.14475","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"SVD-based deconvolution reduced injection-dependent variation in quantitative angiography parameters across patient-specific intracranial aneurysm phantoms.","lead":"This paper tests whether three SVD deconvolution methods can remove variability caused by injection conditions in quantitative angiography of brain aneurysm phantoms. It reports that these methods make measured blood flow parameters more consistent across injection speeds, aneurysm sizes and locations, and stenosed or normal parent arteries.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing ground truth means the reported slope reductions could reflect regularization-induced shrinkage rather than genuine bias reduction; the central claim as stated is not yet established.","rationale":"The reader identified the LTI convolution assumption as the weakest point. That is a real concern, but the more decisive issue is that even granting LTI, the paper supplies no ground truth against which 'bias reduction' can be judged. The measured outcome is only a reduction in slope, which is compatible with genuine deconvolution, with regularization-induced shrinkage, or with the mathematical identity of the deconvolution operation under the assumed model. The absence of multi-geometry, multi-bolus, and statistical test results further limits the generality claimed in the abstract. Because a conditional verdict already requests more evidence, my concern does not change the verdict, but it sharpens what the missing evidence must be: accuracy validation, not just consistency.","tokens_in":6327,"tokens_out":5063,"duration_ms":53624,"concrete_test":"Generate simulated TDCs with known ground-truth IRF and known MTT by convolving a known impulse response with each measured inlet AIF, add noise at the same level as the phantom data, and run the exact sSVD, bSVD, and oSVD implementations. Check whether the recovered MTT values track the true MTT across a range of true transit times and whether the slope-vs-duration reduction persists without systematic underestimation. Additionally, for at least one phantom configuration, compare the deconvolved MTT to an independent measurement, such as ultrasound-derived flow or CFD transit time, to confirm absolute accuracy. If the recovered values cluster near a value unrelated to the true MTT, the central bias-reduction claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that SVD variants 'significantly reduced QA parameter variability due to injection techniques,' and the title and conclusion extend this to 'bias reduction.' What the results actually show is that, after deconvolution, the slope of MTT versus injection duration is smaller for one aneurysm geometry and one bolus volume. Under the assumed LTI convolution model in Section 2.2, this outcome is partly a mathematical consequence: if the aneurysm TDC equals the convolution of the inlet AIF with a fixed impulse response, then deconvolving with the measured inlet TDC should remove injection-duration dependence by construction. The empirical result therefore cannot validate the model or the physiological accuracy of the recovered parameters. In particular, Table 1 reports oSVD non-stenosed MTT_SVD values of about 0.08-0.11 s while the raw TDC MTT values are about 1.9-2.6 s; without an independent reference, one cannot tell whether this is correct deconvolution or an artifact of Tikhonov or oscillation-index regularization combined with 15 fps sampling. A method that systematically shrinks all estimates toward a near-constant value would also produce a smaller slope, so 'bias reduction' as accuracy relative to truth is not established by the presented evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SVD-based deconvolution methods (sSVD, bSVD, oSVD) as a means of reducing hand-injection variability in quantitative angiography (QA) of intracranial aneurysms. In an in-vitro patient-specific phantom setup with three aneurysm geometries at different Circle of Willis locations, the authors acquired DSA time-density curves for non-stenosed and stenosed parent arteries, with 5 ml and 10 ml boluses at four injection rates. They deconvolved the inlet TDC (AIF) from the dome TDC to obtain an impulse response function, from which they computed PH_IRF, AUC_IRF, and MTT. The central claim, stated in the Abstract, is that the SVD variants 'significantly reduced QA parameter variability due to injection techniques.' Detailed results are shown only for aneurysm geometry 1 with a 5 ml bolus, reporting slope reductions of MTT versus injection duration of 92.56% (sSVD), 98.01% (bSVD), and 98.45% (oSVD) for the non-stenosed artery, and corresponding reductions for the stenosed artery. The paper concludes that SVD-based deconvolution is a valid normalization method for QA analysis.","tokens_in":6478,"tokens_out":3476,"duration_ms":36196,"significance":"If the claims were fully supported, this would be a valuable contribution to neurointerventional imaging, because hand-injection variability is a known limitation of QA, and a demonstrably robust deconvolution correction would improve the clinical utility of DSA-derived hemodynamic parameters. The use of patient-specific phantoms with realistic flow waveforms, multiple injection protocols, and stenosed as well as non-stenosed conditions is a strength, and the three SVD variants are well-motivated from the perfusion-imaging literature. However, the evidence presented in the manuscript is substantially narrower than the abstract and conclusion claim: numerical results are given for only one of the three aneurysm geometries and only one bolus volume, no statistical significance tests support the word 'significantly,' and no independent ground truth is available to distinguish genuine bias reduction from regularization-induced shrinkage. The core idea remains plausible, but the current manuscript does not yet establish the claimed generality or the accuracy component of 'bias reduction.'","major_comments":[{"comment":"The central claim of significant variability reduction 'across various aneurysm configurations' is not supported by the presented data. The Results section states that the experiment was repeated with three aneurysm geometries but 'we have provided the results for aneurysm geometry 1 ... having a 5ml bolus' only. No quantitative results, slopes, or parameter tables are shown for the other two geometries or for the 10 ml bolus condition. Since the abstract and conclusion generalize to different aneurysm sizes, locations, and injection volumes, the missing results are load-bearing; the manuscript should either include the full dataset or explicitly limit the claims to the conditions actually reported.","section":"Abstract and Results (Figure 3, Table 1)"},{"comment":"The word 'significantly' in the Abstract and Conclusion is not backed by any statistical test. The table reports means and standard deviations for three repeats, but there is no t-test, ANOVA, confidence interval for the slopes, or comparison of the pre- and post-deconvolution slopes. The percentage slope reductions could be tested against the null hypothesis of no reduction, and the variability across injection rates should be compared in a formal way. As written, 'significantly' has no statistical meaning, and the claim should be reworded or supported with appropriate tests.","section":"Results (Figure 3, Table 1)"},{"comment":"Under the assumed linear time-invariant (LTI) convolution model, the deconvolution step removes injection-duration dependence by construction: if the dome TDC is exactly the convolution of the inlet AIF with a fixed impulse response, then the recovered IRF and its MTT are independent of the input function. The observed slope reduction therefore tests consistency with LTI behavior, not the accuracy of the recovered QA parameters. The large discrepancy between raw MTT values (about 1.9-2.6 s) and oSVD-derived MTT values (about 0.08-0.11 s) in Table 1 is not discussed, and without an independent reference (e.g., a known IRF, flow measurements, or CFD-computed TDCs) it is unclear whether the near-flat slopes reflect true physiological transit times or an artifact of regularization that shrinks IRF estimates toward a near-constant value. The manuscript should address this ambiguity and temper the use of the term 'bias reduction' unless accuracy against a ground truth is demonstrated.","section":"Section 2.2, Eq. (deconvolution model)"},{"comment":"The manuscript does not disclose the Tikhonov regularization parameter used in sSVD, the truncation threshold or other regularization settings for bSVD, or the oscillation-index regularization parameter and penalty strength for oSVD. These settings control the bias-variance tradeoff and directly influence the shape of the recovered IRF and hence MTT, PH_IRF, and AUC_IRF. Without these values, the results are not reproducible, and the reader cannot assess whether the reported slope reductions are sensitive to arbitrarily chosen regularization strengths. The authors should state the exact parameter values and, if possible, show sensitivity of the main results to those parameters.","section":"Section 2.2, Methods (regularization parameters)"}],"minor_comments":[{"comment":"The keyword 'Single Value Decomposition' should be corrected to 'Singular Value Decomposition.'","section":"Keywords"},{"comment":"The phrase 'AUCIRF which is calculated by integrating the IRF with respect to AIF' is ambiguous; it should be 'integrating the IRF over time' or 'with respect to time,' since the area under the IRF curve is a time integral.","section":"Section 2.2"},{"comment":"The reconstructed TDC Qnew is mentioned but never used in the analysis or validation; if Qnew is intended as a quality check of the deconvolution, the authors should report the residuals, correlation, or another fit metric.","section":"Section 2.2 and Figure 2"},{"comment":"The table formatting is difficult to follow: the row and column alignment for DIFFSL and AVGSL is inconsistent, and some cells appear misplaced or missing. The caption also introduces SLOPESVD, SLOPE, DIFFSL, and AVGSL, but the table does not clearly show all of these quantities for every row. A clearer layout or a separate table for slope statistics would improve readability.","section":"Table 1"},{"comment":"The axis labels and symbols in Figure 3 are not fully legible in the provided version; the authors should ensure that the units of MTT are explicitly stated on the y-axis and that stenosed versus non-stenosed curves are clearly distinguished by markers or legends.","section":"Figure 3"},{"comment":"Reference [17], 'A Survey of Singular Value Decomposition Methods for Distributed Tall/Skinny Data,' appears to be a computing-systems reference that is not directly relevant to the medical-imaging context; a standard numerical linear algebra reference (e.g., Golub and Van Loan) would be more appropriate for the SVD description.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for physics.med-ph, and the experimental platform is a genuine strength. However, the gap between the abstract's broad claim and the single-geometry, single-bolus analysis actually presented is substantial. If the authors have the full dataset for all three geometries and both bolus volumes, they should be encouraged to include it; if not, the claims must be narrowed. The missing regularization parameters and the absence of any statistical analysis are additional reasons that this revision should be more than a minor edit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on this one.\n\nWhat's new: they built a patient-specific phantom with three aneurysm geometries at different Circle of Willis locations, stenosed and non-stenosed parent arteries, two bolus volumes, and four injection rates, each repeated three times. That's real experimental effort, and it's the kind of data the QA community needs. Applying sSVD, bSVD, and oSVD to extract IRF-based parameters is not new—perfusion imaging has used these for decades, and the same group's ref [6] already showed SVD reduces injection variability in simpler phantoms—but the patient-specific multi-location setup is a legitimate extension.\n\nThe reported effect is plausible: for geometry 1 with a 5 ml bolus, the slope of MTT versus injection duration drops by 92–98% after deconvolution. I believe the deconvolution is doing what it's supposed to do under the linear time-invariant model they assume.\n\nThe soft spots are in proportion to the strength of the claim. The abstract and title say 'significantly reduced' and 'bias reduction,' but the evidence only shows internal consistency for one geometry and one bolus volume. No significance tests accompany 'significantly.' The sSVD Tikhonov parameter is not reported, so the deconvolution isn't fully reproducible. More importantly, there's no independent ground truth. Under the LTI model, if the dome TDC is truly the convolution of the inlet AIF with a fixed IRF, then deconvolving with the measured inlet TDC removes the injection duration by construction. The empirical slope reduction therefore can't validate the model or prove the recovered PH, AUC, and MTT are accurate. The oSVD non-stenosed MTT values around 0.08–0.11 s are a red flag: they may be real IRF transit times, but without a reference standard you can't rule out regularization shrinking everything toward a near-constant value, which would also flatten the slope. The paper needs a simulation or a phantom with a known IRF to show the deconvolution recovers the true parameters, not just a smaller slope.\n\nThe citation pattern is fine—ref [6] is the prior work and should be cited. No circular derivation.\n\nBottom line: this is a solid empirical contribution in need of revision. The data collection is worth preserving, but the central claim as written is not established. I'd send it to peer review and ask for full multi-geometry results, significance testing, regularization details, and an accuracy check against a known reference. That's a lot of work, but it's the difference between a conference-style report and a paper that actually supports 'bias reduction.'","headline":"Useful phantom dataset and a plausible effect, but the 'bias reduction' claim outruns the reported evidence — no ground truth, one geometry shown, no significance tests.","tokens_in":7109,"tokens_out":2803,"would_cite":false,"duration_ms":27107,"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":"By deconvolving the aneurysm dome time-density curve with the inlet arterial input function, SVD-based methods reduce the injection-induced variability in quantitative angiography parameters, cutting the mean transit time slope by 92–98%…","keywords":["quantitative angiography","intracranial aneurysm","singular value decomposition","deconvolution","injection variability","impulse response function","patient-specific phantom","digital subtraction angiography"],"falsifier":"Repeat the same phantom experiment with bolus rates far outside the tested range (for example, a 2 ml bolus at 20 ml/s and a 20 ml bolus at 2 ml/s) and compare the recovered IRFs: if peak height, area, or mean transit time changes systematically with injection rate, the convolution model is not linear time-invariant and the reported invariance would not generalize beyond the tested conditions.","tokens_in":6083,"feed_emoji":"🧠","tokens_out":7416,"duration_ms":70014,"temperature":0.7,"pith_summary":"The paper is trying to establish that the variability hand injection introduces into quantitative angiography of intracranial aneurysms can be removed by deconvolving the aneurysm's contrast curve with the inlet arterial curve using singular value decomposition (SVD) variants. It uses patient-specific in vitro phantoms with three aneurysm geometries at the middle cerebral, anterior communicating, and internal carotid arteries, under both healthy and 75%-stenosed parent-artery conditions, with 5 ml and 10 ml boluses at four injection rates. For the reported geometry, the slope of mean transit time versus injection duration decreased by 92.56% with standard SVD, 98.01% with block-circulant SVD, and 98.45% with oscillation-index SVD in the non-stenosed case, and by 99.38%, 74.25%, and 75.37% in the stenosed case. If the invariance holds generally, QA parameters such as peak height, area under the curve, and mean transit time could be compared across patients and treatment stages without strict injector standardization.","feed_headline":"SVD deconvolution flattens injection bias in aneurysm imaging","feed_subtitle":"Mean transit time slope drops by up to 98 percent across injection rates in phantom aneurysm studies.","key_machinery":"The central object is the impulse response function recovered by singular value decomposition of the arterial input function matrix. In standard SVD the inlet time-density curve is arranged as a Toeplitz matrix; block-circulant SVD and oscillation-index SVD use a block-circulant matrix, with the oscillation-index variant also computing an oscillation index to regularize the solution. Decomposing the AIF matrix into singular vectors and singular values, and then solving for the IRF, is what strips the injection waveform out of the aneurysm curve. The IRF then carries the hemodynamic information, and the QA parameters are read off its peak, its integral, and the ratio of integral to peak.","core_discovery":"The paper claims that the aneurysm dome time-density curve Q is the convolution of the inlet arterial input function Ca with a single impulse response function (IRF), and that SVD-based deconvolution inverts this relation. After deconvolution, the extracted IRF is convolved back with Ca to produce a reconvolved curve Qnew that closely tracks the measured aneurysm curve. Because the IRF is meant to encode only the aneurysm's transport behavior, its derived parameters—peak height (PHIRF), area under the curve (AUCIRF), and mean transit time (MTT)—become nearly independent of bolus volume and injection duration. The paper reports that the slope of MTT versus injection duration drops almost to zero for aneurysm geometry 1 with a 5 ml bolus, and presents this invariance as evidence that the SVD variants standardize quantitative angiography across aneurysm sizes, locations, and parent-artery conditions.","pith_inferences":["Editorial inference: The linear time-invariant assumption has a testable consequence the paper does not state—if the recovered IRF is truly injection-independent, it should also remain unchanged under variations in pump stroke volume or heart rate, which alter the flow state rather than the injection.","Editorial inference: A practical downstream gain is that SVD-normalized QA parameters could be pooled across clinical sites or compared across DSA systems that use different injector settings, since deconvolution is meant to absorb the injection waveform.","Editorial inference: The different behavior in the stenosed artery, where standard SVD outperformed the other variants, hints that residual slope after deconvolution might itself encode information about the parent-artery flow condition; a future study could test whether the post-SVD slope separates stenosis severity from injection effects."],"forward_implications":["If SVD-deconvolved QA parameters are truly invariant to injection conditions, serial DSA runs in the same patient can be compared without matching injection volume or rate, making device-induced hemodynamic changes easier to track over time.","The near-zero MTT slopes reported for the non-stenosed artery mean that after correction, the same aneurysm yields essentially the same transit time whether contrast is delivered over a quarter second or a full second.","The fact that the correction also reduces MTT slope in the stenosed parent artery implies the method does not require normal flow to work, which matters for the diseased vessels where treatment decisions are made.","Because the three SVD variants reduce variability by different amounts in the main non-stenosed case, the choice of variant changes the tightness of the resulting parameter invariance.","The paper reports robustness across aneurysm sizes and locations, which would make the correction a general normalization step rather than a geometry-specific fix."],"supporting_citations":[{"why":"Prior phantom study establishing that SVD can remove injection variability in 2D quantitative angiography, which this paper extends to multiple aneurysm geometries and stenosed arteries.","marker":"[6]"},{"why":"Phantom study of injection technique effects on temporal parametric imaging that motivates the experimental setup and the need for correction.","marker":"[16]"},{"why":"Introduces the block-circulant deconvolution matrix used by bSVD to make the method insensitive to tracer arrival timing.","marker":"[12]"},{"why":"Supplies the Tikhonov regularization used in the sSVD variant to stabilize the deconvolution.","marker":"[10]"},{"why":"Shows truncated SVD applied to DSA-derived parametric imaging maps, providing the SVD-on-DSA basis the QA pipeline builds on.","marker":"[8]"},{"why":"Supplies the singular-value properties of block circulant matrices that underlie the bSVD and oSVD construction.","marker":"[11]"}],"fun_headline_variants":["SVD kills injection bias in aneurysm angiography","Deconvolution tames hand-injection error in aneurysm imaging","Bolus size no longer matters in aneurysm imaging with SVD","Injection variability erased by SVD in aneurysm scans","Phantom study proves SVD standardizes aneurysm quantification"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The correction works only if one linear, time-invariant impulse response function fully describes contrast transport through the aneurysm regardless of injection rate, volume, or flow state, and only if the inlet time-density curve faithfully represents the contrast actually entering the aneurysm.","fun_headline_variants_meta":{"raw":{"variants":["SVD kills injection bias in aneurysm angiography","Deconvolution tames hand-injection error in aneurysm imaging","Bolus size no longer matters in aneurysm imaging with SVD","Injection variability erased by SVD in aneurysm scans","Phantom study proves SVD standardizes aneurysm quantification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00071,"raw_usage":{"total_tokens":3247,"prompt_tokens":1047,"completion_tokens":2200,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":663,"completion_tokens_details":{"reasoning_tokens":2122}},"tokens_in":663,"tokens_out":2200,"duration_ms":15998,"temperature":1.0,"reasoning_tokens":2122,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:50:30.392957+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the same phantom experiment with bolus rates far outside the tested range (for example, a 2 ml bolus at 20 ml/s and a 20 ml bolus at 2 ml/s) and compare the recovered IRFs: if peak height, area, or mean transit time changes systematically with injection rate, the convolution model is not linear time-invariant and the reported invariance would not generalize beyond the tested conditions.","supporting_citations":[{"cited_title":"Effect of singular value decomposition on removing injection variability in 2D quantitative angiography: An in silico and in vitro phantoms study,","cited_arxiv_id":null,"evidence_quote":"Prior phantom study establishing that SVD can remove injection variability in 2D quantitative angiography, which this paper extends to multiple aneurysm geometries and stenosed arteries."},{"cited_title":"Effect of inject ion technique on temporal parametric imaging derived from digital subtraction angiography in patient specific phantoms,","cited_arxiv_id":null,"evidence_quote":"Phantom study of injection technique effects on temporal parametric imaging that motivates the experimental setup and the need for correction."},{"cited_title":"Tracer ar rival timing -insensitive technique for estimating flow in MR perfusion -weighted imaging using singular value decomposition with a block-circulant deconvolution matrix,","cited_arxiv_id":null,"evidence_quote":"Introduces the block-circulant deconvolution matrix used by bSVD to make the method insensitive to tracer arrival timing."},{"cited_title":"On Tikhonov regularization, bias and variance in nonlinear system identification,","cited_arxiv_id":null,"evidence_quote":"Supplies the Tikhonov regularization used in the sSVD variant to stabilize the deconvolution."},{"cited_title":"Effect of Truncated Singular Value Decomposition on Digital Subtraction Angiography Derived Angiographic Parametric Imaging Maps,","cited_arxiv_id":null,"evidence_quote":"Shows truncated SVD applied to DSA-derived parametric imaging maps, providing the SVD-on-DSA basis the QA pipeline builds on."},{"cited_title":"On singular values of block circulant matrices,","cited_arxiv_id":null,"evidence_quote":"Supplies the singular-value properties of block circulant matrices that underlie the bSVD and oSVD construction."}],"review_version":1}