{"id":"5d533a09-be89-432b-832c-a19e26015ef3","arxiv_id":"2502.02728","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A gamma-variate fit to 4DCT contrast enhancement produces voxel-wise maps of arrival time and residence time across the left atrial appendage.","lead":"Using dynamic 4D CT images taken during a contrast injection, the authors fit a gamma-variate curve to each pixel's contrast signal and generate maps of arrival time and residence time across the left atrial appendage. These maps are intended to identify slow-flow regions that could predict blood clot and stroke risk in atrial fibrillation patients, offering a CT-only alternative to echo or MRI flow measurement.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Quantitative RT and arrival-time maps rest on unvalidated gamma-variate fits to 27 sparse, possibly truncated samples; the derived residence time depends on an extrapolated tail that no data may constrain.","rationale":"The reader's conditional verdict is appropriate, and my concern reinforces it rather than changing it. The reader identified baseline and registration as the weakest assumption; I would sharpen the emphasis to the gamma-variate fit and truncation of the washout tail, because Eq. (2) integrates the analytic model and therefore inherits any extrapolation error. The paper's own Discussion notes that the number and timing of images needed for reasonable parameter estimates remains an open question, and no reference-standard comparison is reported. This is not an internal inconsistency, but it is a correctness risk for the central quantitative claim. If the proposed simulation reveals large truncation-related bias, the verdict should move to REJECT; until that test is run, conditional remains the correct verdict.","tokens_in":5100,"tokens_out":5131,"duration_ms":54697,"concrete_test":"Generate synthetic LAA contrast curves with known arrival time and RT from a compartment or CFD model, sample them at the exact 27 acquisition timepoints of the protocol, add noise at the measured signal level, apply the identical pipeline (aortic baseline subtraction, Gaussian smoothing, Table I bounds, gamma-variate fit, Eq. 2), and compare estimated versus true RT and arrival time as a function of true RT. If the relative error or bias grows when true RT approaches or exceeds the last sampled washout time (for example, >20% error or a strong correlation between error and true RT), the claim that RT maps are quantitative is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that voxel-wise gamma-variate fits of 4DCT contrast curves yield quantitative flow parameters (arrival time, RT) in the LAA. The load-bearing, least-secure condition is that the fitted curve, not the raw data, determines these parameters, and that curve is not checked against any independent flow measurement. In Section II.D, after subtracting a baseline computed as the mean of timepoints 1-3 in the descending aorta, each voxel's 27-point signal is fit with Eq. (1) using the bounds of Table I. RT is then computed by Eq. (2), an integral over the analytic gamma-variate. Because 26 volumes are collected over 40-60 s plus one at ~100 s, any voxel whose washout is not complete within the sampled window has a tail constrained only by the model's exponential decay, not by data. In slow-filling/stasis regions, precisely the target of the method, the fitted RT can be an extrapolation. The aortic baseline is also not shown to equal the LAA's true pre-contrast voxel value; a constant offset error shifts the fitted curve and biases t_arrival and RT. The paper reports no comparison with 4D-flow MRI, TEE LAA emptying velocity, CFD, or an injected-contrast phantom, so internal smoothness of the maps is the only evidence that the parameters are physical. This is an externally unvalidated assumption, not an internal inconsistency, but it is load-bearing for the paper's quantitative claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a method to estimate blood flow parameters in the left atrial appendage (LAA) from dynamic contrast-enhanced 4DCT. A 4DCT acquisition protocol with 27 timeframes is used, followed by nonrigid registration and multiplanar reconstruction. Each voxel's time-intensity curve is fit with a gamma-variate model, and parameters including arrival time and residence time (RT) are derived from the fitted curve to create 2D parameter maps. The authors show qualitative results for two patients and state that the maps enable quantification and visualization of flow parameters across the LAA, with potential implications for stroke risk assessment in atrial fibrillation patients.","tokens_in":5429,"tokens_out":2624,"duration_ms":27020,"significance":"If validated, the method could provide a clinically accessible, high-spatial-resolution surrogate for blood stasis in the LAA from routine CT data, which would be valuable for stroke risk stratification. The paper demonstrates a clear technical pipeline and the use of a widely accepted indicator-dilution model. The maps shown appear visually plausible, and the acquisition protocol uses a radiation dose comparable to standard cardiac CT. However, the current evidence is preliminary: no independent validation against TEE, 4D flow MRI, CFD, or a phantom is presented, and the quantitative claim rests on an unvalidated baseline subtraction and extrapolated gamma-variate tails. The cohort of 17 patients is mentioned but results are only shown for 2, with no quantitative summary or error analysis.","major_comments":[{"comment":"Residence time is computed as an integral over the fitted gamma-variate curve using Eq. (2). The 27 timepoints are acquired over 40–60 s plus one at ~100 s, so for voxels with slow washout—precisely the stasis regions of interest—the tail of the curve is extrapolated and not constrained by measured data. This makes the derived RT potentially an artifact of the model's exponential decay rather than a measured physiological quantity. Please provide per-voxel fit uncertainty estimates (e.g., confidence intervals from residuals) or restrict the analysis to voxels where the washout portion is actually sampled.","section":"Section II.D and Eq. (2)"},{"comment":"The baseline y0 is computed as the mean of timepoints 1–3 in a descending aorta ROI. This assumes that the LAA voxel's pre-contrast CT intensity is identical to the aortic baseline, which is not validated. A constant offset error in y0 will bias the fitted y_max, t_arrival, and RT. The authors should validate this assumption by comparing the pre-contrast LAA voxel values to the aortic baseline, or by using a phantom with known contrast concentrations to establish the accuracy of the baseline correction.","section":"Section II.D, baseline subtraction"},{"comment":"The paper reports that 17 patients underwent imaging, but quantitative results are shown for only 2 patients. There is no summary of the range of RT values, inter-patient variability, or the reproducibility of the maps, nor any comparison with independent flow measurements (e.g., TEE LAA emptying velocity, 4D flow MRI, or CFD). The claim that the method 'enables quantification and visualization of spatial-temporal characteristics of flow parameters' is therefore not supported by the evidence presented. Please include a quantitative validation study or a statistically meaningful analysis across the full cohort, and compare the derived parameters against an independent reference.","section":"Section III.B and Fig. 4"},{"comment":"The gamma-variate fit uses loose parameter bounds (alpha unbounded, y_max and t_arrival within ±40 HU and ±5 s of the data maximum). No assessment of fit quality (e.g., R^2, chi-square, or residual analysis) is reported per voxel or per patient. It is unclear whether the smooth maps reflect physiological variation or simply the smoothness of the fitting function. Please report fit quality metrics and consider adding a goodness-of-fit criterion to exclude poorly fitted voxels.","section":"Table I and Section II.D"}],"minor_comments":[{"comment":"The word '4-dimentional' should be 'four-dimensional'.","section":"Abstract"},{"comment":"The threshold of cumulative sum <100 HU for voxel exclusion is introduced without justification; please explain how this value was chosen and its sensitivity.","section":"Section II.B"},{"comment":"The parameter t_arrival is not clearly introduced before it is used; please define it explicitly and consistently (the text appears to have formatting artifacts for the symbol).","section":"Section II.D"},{"comment":"The term '2D fit over time' is ambiguous; please clarify whether this refers to fitting the temporal signal at each pixel or to a spatial 2D fit.","section":"Fig. 1"},{"comment":"In reference [10], a comma is missing after 'Z.M. Lin'; please verify the reference formatting.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The core idea is clinically relevant and technically feasible, but the paper's quantitative claims outrun its evidence. The lack of any independent validation is the main obstacle; the manuscript would be significantly strengthened by adding a comparison to an established flow measurement or a phantom study. The limited number of displayed patients (2 of 17) and the absence of error analysis further weaken the current version. I recommend major revision rather than rejection because the method is novel and the identified issues are addressable within the scope of the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a feasibility report, not a validated quantitative method. The genuinely new piece is applying standard gamma-variate DCE analysis to LAA voxels from 4DCT, yielding smooth maps of arrival and residence times in two illustrative patients. The imaging protocol is described clearly, including the low-dose 100 mA setting, nonrigid registration, and 27 timepoints, and the use of an established perfusion model is appropriate. Credit is due for transparency about the pipeline and for distinguishing direct fit parameters from derived ones.\n\nThe soft spots are real and load-bearing. RT is computed by integrating the analytic fitted curve, so for voxels whose washout is incomplete within the 40-60 s sampling window—exactly the slow-flow regions that matter—the tail is pure model extrapolation. The baseline is taken from the descending aorta at timepoints 1-3, with no evidence that it matches the LAA's pre-contrast signal; any offset biases every derived parameter. Only two selected patients are shown; there are no error bars, no sensitivity analysis for the 100 HU threshold or Gaussian smoothing, and no comparison with TEE, 4D-flow MRI, or CFD. The claim that these maps \"quantify\" flow is therefore unsupported.\n\nI don't see circularity: the parameters are definitional functions of the fitted curve, not the result of refitting the target conclusion. The weakness is external validation, not internal logic. The stress-test worry about the extrapolated tail is the strongest point, but the aortic-baseline assumption is almost as serious.\n\nWho is this for? Readers working on CT-based stasis imaging or LAA thrombus risk. It is a reasonable pilot with sound clinical motivation, and a serious referee could help the authors either add validation or reframe the paper as a feasibility study and soften \\\"quantification\\\".\n\nRecommendation: send it to peer review, but the realistic first decision is major revision, not acceptance.","headline":"A feasible CT-only pipeline for mapping LAA contrast dynamics, but the quantitative flow claims rest on unvalidated gamma-variate fits and two illustrative cases.","tokens_in":5946,"tokens_out":1773,"would_cite":false,"duration_ms":19072,"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":"Fit each 4DCT voxel's contrast curve with a gamma-variate model and the result is a spatial map of blood arrival and residence time in the left atrial appendage.","keywords":["dynamic contrast enhancement","4DCT","left atrial appendage","gamma-variate fit","residence time","atrial fibrillation","blood stasis","CT perfusion"],"falsifier":"Compare the residence-time maps against LAA flow velocity measured by transesophageal echocardiography or 4D-flow MRI in the same atrial-fibrillation patients: if regions flagged as high residence time show fast contrast washout on those modalities, the maps do not represent stasis. A simpler mechanical check is a flow phantom with known chamber transit times imaged with the same 27-timeframe protocol, requiring the fitted residence times to match the known values within tolerance.","tokens_in":4904,"feed_emoji":"🫀","tokens_out":5450,"duration_ms":44863,"temperature":0.7,"pith_summary":"The paper proposes that the time-varying CT signal from a single contrast-enhanced 4DCT scan can be read as a local indicator-dilution curve throughout the left atrial appendage (LAA). At every voxel, the 27 timepoints of contrast enhancement are fit with a gamma-variate curve after nonrigid motion correction, and the fitted curve yields two flow parameters: arrival time (when signal reaches 1% of its peak) and residence time (the intensity-weighted mean transit time). The fitted parameters are presented as smooth two-dimensional maps that show where contrast arrives late and lingers, which the authors interpret as a direct visualization of blood stasis in atrial fibrillation patients. If correct, this gives clinicians a way to assess stroke risk from an ECG-gated CT already performed for cardiac morphology, without solving fluid-dynamics equations. The claim is that these maps capture spatial-temporal flow characteristics across the LAA, not that CT measures flow directly.","feed_headline":"Gamma-variate fits turn 4DCT contrast into LAA blood-flow maps","feed_subtitle":"Per-voxel contrast curves become smooth residence-time maps, a potential stasis marker in atrial fibrillation.","key_machinery":"The load-bearing object is the gamma-variate model, a three-parameter curve widely used to describe the passage of an injected indicator bolus through a series of compartments. Written as $y(t) = y_{\\max}\\,(\\alpha e)^{-\\alpha}(t/t_{\\mathrm{peak}})^\\alpha - y_b$, it turns each voxel's noisy 27-point contrast time series into a smooth analytical curve. The fitted curve supplies direct parameters ($y_{\\max}$, $t_{\\mathrm{peak}}$, $\\alpha$) and derived transit parameters ($t_a$ and $RT$) that become the values plotted on dynamic contrast enhancement maps. The model's role is to regularize the sparse, noisy temporal signal so that per-voxel flow parameters can be mapped without averaging away spatial detail.","core_discovery":"On the paper's own terms, the discovery is that voxel-wise gamma-variate fitting of 4DCT dynamic contrast enhancement produces clinically interpretable maps of LAA blood flow parameters. For each voxel, the contrast signal is baseline-corrected using the mean of the first three timepoints in the descending aorta, then fit to the gamma-variate function $y(t) = y_{\\max}\\,(\\alpha e)^{-\\alpha}(t/t_{\\mathrm{peak}})^\\alpha - y_b$, with fit parameters $\\alpha$, $y_{\\max}$, and $t_{\\mathrm{peak}}$. From the fitted curve the authors derive arrival time $t_a$, defined as the time the signal reaches 1% of its peak, and residence time $RT = \\int t\\,y(t)\\,dt / \\int y(t)\\,dt$. Applied across a multiplanar slice through the LAA, these parameters form smooth maps, and in two illustrative patients the maps separate a uniform-filling phenotype from a delayed-filling phenotype in the middle and distal LAA. The paper frames this as enabling quantification and visualization of spatial-temporal flow characteristics and, potentially, thrombus risk.","pith_inferences":["A natural next step, not stated in the paper, is to use the ratio $RT / t_a$ or the peak-normalized curve width as a dimensionless stasis index that could be compared across patients independent of heart rate and injection timing.","The same per-voxel gamma-variate pipeline applied to the full left atrial volume, rather than a single 2D slice, would produce 3D stasis maps; the authors note they are working toward this with a machine-learning approach.","The maps' clinical value hinges on an external anchor: comparing residence-time maps with LAA emptying velocity from transesophageal echocardiography or 4D-flow MRI in the same patients would test whether 'slow filling' on CT matches 'stasis' by established measures.","If the gamma-variate residuals are structured (e.g., high near the LAA wall), the maps may be capturing a mixture of motion-correction error and true slow flow; inspecting residual maps would separate these."],"forward_implications":["A single ECG-gated 4DCT acquisition during a standard contrast injection can yield voxel-resolved maps of arrival time and residence time across the LAA, with no CFD simulation required.","Maps that are uniform across the LAA (as in Patient 1) and maps with delayed distal filling (as in Patient 2) give a direct visual readout of filling patterns that may correspond to differing thrombus risk.","Because the mean radiation dose is comparable to other cardiac CT exams, the approach could be added to existing clinical CT protocols for AF patients without a separate imaging study.","The success of fits from 27 sparse timepoints suggests the temporal sampling rate could be reduced, with the dose lowered further if fewer timeframes prove adequate."],"supporting_citations":[{"why":"Establishes the gamma-variate as the standard model for indicator transit time, the conceptual basis for interpreting contrast curves as flow.","marker":"[8]"},{"why":"Provides the simplified gamma-variate formulation the paper uses for fitting.","marker":"[9]"},{"why":"Supplies the specific gamma-variate fitting approach for quantitative CT perfusion that this LAA application adapts.","marker":"[11]"},{"why":"Supports gamma-variate fitting over deconvolution for CT perfusion curves.","marker":"[10]"},{"why":"Describes the dynamic CT myocardial perfusion protocol on which the acquisition sequence is based.","marker":"[6]"},{"why":"Provides the nonrigid 4D motion correction that aligns all timeframes before fitting.","marker":"[7]"},{"why":"Demonstrates 4D-flow MRI as an established modality for LAA flow dynamics, the standard this CT surrogate would need to match.","marker":"[5]"}],"fun_headline_variants":["4DCT gamma-variate maps expose LAA stasis","Gamma-variate fits convert 4DCT to LAA flow maps","Residence-time maps from 4DCT show LAA flow","4DCT contrast curves pinpoint LAA slow flow","Voxel gamma fits map LAA stasis from 4DCT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The premise the method depends on is that after motion correction and subtraction of a descending-aorta baseline, each voxel's CT intensity over time tracks the local contrast-agent concentration faithfully enough that the gamma-variate fit returns true arrival and residence times for LAA blood.","fun_headline_variants_meta":{"raw":{"variants":["4DCT gamma-variate maps expose LAA stasis","Gamma-variate fits convert 4DCT to LAA flow maps","Residence-time maps from 4DCT show LAA flow","4DCT contrast curves pinpoint LAA slow flow","Voxel gamma fits map LAA stasis from 4DCT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000706,"raw_usage":{"total_tokens":3208,"prompt_tokens":997,"completion_tokens":2211,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":2121}},"tokens_in":613,"tokens_out":2211,"duration_ms":14611,"temperature":1.0,"reasoning_tokens":2121,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T11:20:37.991714+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the residence-time maps against LAA flow velocity measured by transesophageal echocardiography or 4D-flow MRI in the same atrial-fibrillation patients: if regions flagged as high residence time show fast contrast washout on those modalities, the maps do not represent stasis. A simpler mechanical check is a flow phantom with known chamber transit times imaged with the same 27-timeframe protocol, requiring the fitted residence times to match the known values within tolerance.","supporting_citations":[{"cited_title":"Comparison of Blood Flow Models and Acquisitions for Quantitative Myocardial Perfusion Estimation from Dynamic CT,","cited_arxiv_id":null,"evidence_quote":"Supplies the specific gamma-variate fitting approach for quantitative CT perfusion that this LAA application adapts."},{"cited_title":"CT perfusion: comparison of gamma-variate fit and deconvolution,","cited_arxiv_id":null,"evidence_quote":"Supports gamma-variate fitting over deconvolution for CT perfusion curves."},{"cited_title":"GPNLPerf: Robust 4d Non-rigid Motion Correction for Myocardial Perfusion Analysis,","cited_arxiv_id":null,"evidence_quote":"Provides the nonrigid 4D motion correction that aligns all timeframes before fitting."}],"review_version":1}