{"id":"7c37d79a-41d4-4394-a727-e2eb880b0615","arxiv_id":"2501.10615","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A hierarchical Bayesian LoG stream added to a 3D U-Net improves aorta and supra-aortic branch segmentation and provides confidence intervals.","lead":"This paper introduces LoGB-Net, a two-stream deep learning model that adds a trainable, Bayesian Laplacian-of-Gaussian filter bank to a 3D U-Net to improve segmentation of the aorta and its branching vessels from CT scans. It reports Dice score gains of about 3% over strong baselines and supplies uncertainty intervals meant for downstream blood-flow simulations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Bayesian/UQ component is the least secure load-bearing part: no ELBO or posterior sampling is specified and the confidence intervals are never calibration-tested, leaving the 'reliable confidence intervals' claim unverifiable.","rationale":"I read the paper as making a separable two-part claim: a segmentation accuracy gain and a Bayesian uncertainty-quantification gain. The accuracy part is supported by a reasonably structured comparison (five seeds, standard deviations, multiple baselines, ablations), and the reported margins on SA Dice are internally consistent with the ablation table, so I would not reject the method on those grounds. The least secure load-bearing condition is the UQ claim. The paper offers no actual ELBO expression, no variational family, no posterior-predictive sampling rule, and no calibration/coverage evaluation; the phrase 'ten Bayesian inferences' is the only operational hint. Without these, the title's 'Bayesian neural network' and the abstract's 'reliable confidence intervals' cannot be verified, regardless of whether the queue capacity C is later supplied. The Reader's weak-assumption pick (balanced gate) is real but secondary, since the code repository can resolve it; the Bayesian construction is the place where the central argument is least secure. This does not change the overall CONDITIONAL verdict: the concerns are addressable with a derivation and a coverage experiment, so I would keep the paper conditional rather than accept or reject it now.","tokens_in":8294,"tokens_out":7352,"duration_ms":78308,"concrete_test":"Reconstruct the implemented Bayesian procedure from Section 3 and the released repository. Specifically: (1) write the full ELBO expression actually optimized, including the variational distribution and any KL term; (2) state how the per-voxel confidence intervals are generated from 'ten Bayesian inferences' (weight sampling, MC dropout, or ensemble); (3) run the released code on a held-out AVT test volume and compute the empirical 95% coverage of ground-truth voxels, stratified by SA/MA and by blur/clear boundary regions. If no valid ELBO can be written, the code performs no posterior sampling, or observed coverage departs substantially from 95%, then the 'reliable confidence intervals' portion of the central claim fails even if the Dice gains reproduce.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim pairs a segmentation gain with Bayesian confidence intervals. Table 1 gives a plausible, five-seed comparison for the gain, but the Bayesian construction in Section 3 does not support the UQ claim. The paper writes p(θ|D)=p(D|θ)p(θ)/p(D) and says θ are the LoG parameters, then states θ=argmax(ELBO) and L=Dice−ELBO, but no ELBO is defined: there is no variational family, no KL term, and no posterior predictive. The sentence 'treating the LoG kernel as the prior distribution and the module's output as the posterior' conflates a kernel initialization with a distribution over parameters. The confidence intervals are said to come from 'ten Bayesian inferences,' but the sampling procedure is never described, and the only evaluations are visual (Figure 2) and a CFD illustration. No coverage or calibration metric is reported on held-out volumes, so 'reliable confidence intervals' is asserted rather than demonstrated. The balanced-gate omission (capacity C and inference-time queueing) is a genuine reproducibility gap, but it is secondary: the gate can be pinned down from the released code, whereas the Bayesian/UQ pipeline is currently too underspecified to check at all.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LoGB-Net, a dual-stream 3D aorta segmentation architecture consisting of a regular 3D U-Net stream and a hierarchical Laplacian-of-Gaussian (LoG) stream with a so-called Bayesian parameterization and a balanced gate for foreground-background rebalancing. The method is evaluated on two CTA datasets against ten baselines, with a five-seed comparison reporting Dice, ASD, and Hausdorff distances. The authors claim that LoGB-Net outperforms all baselines, especially on supra-aortic branches (about 3 percentage points higher Dice than Swin-UNETR), and that the Bayesian LoG module provides confidence intervals for segmentation, illustrated on boundary-clarity variations and used to propagate uncertainty into CFD simulations.","tokens_in":8544,"tokens_out":3051,"duration_ms":30429,"significance":"If the claims hold, the paper addresses a genuine clinical need: automated, accurate segmentation of small aortic branches and uncertainty quantification for downstream hemodynamic simulation. The experimental setup is a real, multi-seed comparison against a wide range of modern baselines on two public datasets, and the code is released, which are clear strengths. The ablation study (Table 2) gives initial evidence that each proposed component contributes to the reported performance. However, the Bayesian/UQ component, which is load-bearing for the paper's central 'reliable confidence intervals' claim, is under-specified: no variational objective, inference procedure, or calibration evaluation is provided, and the balanced gate's capacity and inference-time behavior are not defined. The paper's contribution is therefore only partially supported at present.","major_comments":[{"comment":"The ELBO is named but never defined. The text writes 'θ = argmax(ELBO)' and 'L = Dice − ELBO', but no variational family, no KL divergence term, and no posterior predictive distribution are given. The sentence 'treating the LoG kernel as the prior distribution and the module's output as the posterior' appears to conflate a kernel initialization with a distribution over parameters. This makes the Bayesian claim unverifiable and the confidence intervals in Section 4 unsupported. Please provide the full variational objective, the exact inference procedure (e.g., Monte Carlo dropout, mean-field variational inference, or sampling from an approximate posterior), and the mechanism by which the 'ten Bayesian inferences' in the UQ paragraph are generated.","section":"Section 3, 'Bayesian method and balanced gate'"},{"comment":"The queue capacity C is never specified, and the inference-time behavior of the gate is undefined. The text says that when counters c1 and c2 reach capacity C, the 2C images are concatenated along the channel dimension and the regular stream is initiated. At test time, volumes are processed individually; it is unclear whether the gate buffers images across the test set or is disabled during inference. This is load-bearing because Table 2 shows that removing the gate reduces SA Dice from 0.927 to 0.896. Please specify the value of C, the training dynamics when the queues are not full, and precisely how a single test volume is handled. If the gate is not applied at inference, the reported gains may be an artifact of batch-level rebalancing rather than a deployable architectural feature.","section":"Section 3, 'Balanced gate'"},{"comment":"The claim of 'reliable confidence intervals' is asserted but not demonstrated. The confidence intervals are said to come from 'ten Bayesian inferences', yet no sampling method, posterior approximation, or number of stochastic forward passes is described. The only evaluations are visual (Figure 2) and a CFD illustration (Figure 4), which do not constitute calibration evidence. Please report a coverage or calibration metric on held-out volumes, such as the fraction of ground-truth voxels falling within the confidence interval or an expected calibration error, and specify exactly how the ten inferences are generated.","section":"Section 4, 'UQ' and abstract"},{"comment":"The abstract claims 'at least a 3% gain in the Dice coefficient over state-of-the-art methods across multiple volumes.' Table 1 shows a 3.1 percentage-point SA Dice gain over Swin-UNETR, but the MA gain over Swin-UNETR is 2.8 points (0.937 vs. 0.909), and the gain over UNETR++ on MA is 3.0 points (0.937 vs. 0.907). The 'at least 3%' claim is therefore not uniformly true across regions and baselines. Please qualify the claim (e.g., 'up to 3%' or specify the best baseline for each region) so that it matches the reported numbers.","section":"Abstract and Section 4, Table 1"}],"minor_comments":[{"comment":"The numbered pipeline steps are not fully consistent with the preceding description: step (2) says the gate 'initiate[s] the regular stream', while steps (3)-(4) describe concatenating the regular and LoG stream outputs into an ASPP module. Please clarify the exact data flow, including where the LoG stream's mean output enters the concatenation.","section":"Section 3, 'Algorithm Description'"},{"comment":"The equation for T = Σ f_i / Σ b_j does not define what f_i and b_j are (foreground/background voxel counts? intensities?), and the text says the threshold μ is 'determined statistically by calculating the mean of the foreground-to-background ratio for the SA' but does not say over which training images or how the value 0.15 was obtained. Please provide a precise definition and the computation procedure, and consider reporting sensitivity to μ.","section":"Section 3, 'Hierarchical LoG module'"},{"comment":"The caption states that the Bayesian segmentation results' blue, red, and green lines represent the upper bound, mean, and lower bound, but the figure itself does not label which color is which in the displayed image. Please add a legend or clarify in the text.","section":"Figure 1 caption"},{"comment":"The row label 'L−LoG(L(0))' is slightly confusing; it would be clearer to state explicitly that L(0) denotes the regular-stream-only baseline without any LoG stream, and to order the columns consistently with the text describing the ablation.","section":"Table 2"},{"comment":"The UQ-related related work is thin, citing only one variational-inference segmentation paper. Please include representative probabilistic segmentation methods (e.g., PHiSeg, probabilistic U-Net, or MC-dropout approaches) to better position the contribution against the existing UQ literature.","section":"Section 2, 'Related Work'"}],"recommendation":"major_revision","confidential_remarks":"The paper's central risk is that the Bayesian/UQ component is presented as a method but the manuscript does not contain the mathematical or procedural content needed to implement or validate it. The balanced-gate reproducibility gap is smaller but still load-bearing given the ablation results. If the authors can supply the missing ELBO derivation, inference description, calibration experiments, and gate specification, the paper could be publishable. The release of code is helpful, but the manuscript itself must be self-contained for reviewers and readers. Also note that μ=0.15 appears to be chosen from training data; a sensitivity analysis or a principled justification would strengthen the work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know: this is a legitimate empirical paper, not a shell. The authors combine a regular 3D U-Net with a hierarchical trainable LoG stream, add a balanced gate, and report a ~3% Dice gain over ten baselines on supra-aortic branch segmentation, with five random seeds and two public datasets. The ablation shows each component contributes, and the code is linked. The segmentation result itself is plausible and should be taken seriously.\n\nWhat is actually new: the specific design of a five-scale trainable LoG stream with a foreground/background balancing gate for aorta segmentation. That is a modest but useful contribution. The experimental comparison is also better than the field average: ten baselines, consistent training, standard deviations reported.\n\nThe soft spots are real and clustered in the Bayesian part, which the stress-test note names correctly. Section 3 writes a Bayes rule, says theta = argmax(ELBO), and calls the LoG kernel a prior and the module output a posterior, but no ELBO is ever defined: no variational family, no KL term, no posterior predictive, no sampling procedure for the \"ten Bayesian inferences.\" The confidence intervals are shown visually and used in a CFD illustration, but never checked against held-out annotations for coverage or calibration. So the \"reliable confidence intervals\" claim is currently asserted, not demonstrated. That is the load-bearing weak point because it is the paper's main novelty beyond the architecture.\n\nThe balanced gate is under-specified too: queue capacity C is never given, and inference-time behavior is unclear since images are concatenated in groups of 2C during training but inference is presumably per-volume. This is a genuine reproducibility gap, but smaller than the UQ one: the released code could pin it down. The abstract's \"at least a 3% gain\" also overreaches slightly: the gain over Swin-UNETR is about 3.1% on SA but 2.8% on MA, and no significance test backs the \"at least\" phrasing.\n\nBottom line: the architecture and main segmentation result probably hold up. The Bayesian/UQ story needs a full rewrite: specify the inference procedure, report calibration, and temper the claims until then. The gate needs its capacity and inference protocol stated. None of this is fatal; all of it is addressable.\n\nRecommendation: yes, send it to peer review. A serious referee can force the authors to produce the missing derivations and calibration experiments. The paper deserves referee time, but it should not be accepted in its current form.\n\nBest,\n[You]","headline":"Solid, honest architecture paper with real gains on aorta branch segmentation, but the Bayesian/UQ component is presently under-specified and the balanced gate is not fully reproducible; worth refereeing, not ready as is.","tokens_in":9080,"tokens_out":1649,"would_cite":true,"duration_ms":20355,"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":"The paper claims that adding a Bayesian hierarchical Laplacian-of-Gaussian stream to a 3D U-Net lifts aorta and supra-aortic branch segmentation by about 3 Dice points over ten baselines and yields confidence intervals for downstream CFD.","keywords":["aorta segmentation","Laplacian of Gaussian","Bayesian neural network","uncertainty quantification","multiscale vessel segmentation","supra-aortic branches","3D U-Net","CT angiography"],"falsifier":"Force the balanced gate to open on every image by setting its queue capacity to one, retrain on the same 34 volumes, and measure SA Dice on the same 16 test volumes; if the score does not fall from 0.927 toward the 0.896 gate-off level, the gate is not carrying the claimed gain.","tokens_in":8066,"feed_emoji":"🫀","tokens_out":12141,"duration_ms":110275,"temperature":0.7,"pith_summary":"The paper tries to establish that LoGB-Net, a two-stream segmentation model for aortic CT angiography volumes, beats ten existing methods on both the main aorta and the small supra-aortic branches, with the clearest margin on the branches. The first stream is a standard 3D U-Net that captures the whole vessel, and the second stream is a set of five convolutional layers initialized as 3D Laplacian-of-Gaussian kernels at different scales and then trained, which lets the network adapt to vessels of very different radii. A balanced gate rebalances the extreme foreground-to-background ratio, and a Bayesian parameterization of the LoG stream produces per-voxel confidence intervals instead of a single surface. If the claims hold, automated aorta branch segmentation becomes more accurate and downstream hemodynamic simulations get a quantified uncertainty from the segmentation itself.","feed_headline":"Aorta branch Dice jumps ~3 points with a Bayesian LoG stream","feed_subtitle":"A 3D U-Net plus a trainable multiscale edge filter finds small aortic branches and reports confidence intervals.","key_machinery":"The load-bearing object is the hierarchical LoG stream: five 3D convolutional layers of kernel sizes 3, 5, 7, 9, and 11, initialized with discretized 3D Laplacian-of-Gaussian kernels at $\\sigma$ values 0.5, 1.0, 1.5, 2.0, and 2.5, and left trainable. This gives the model a self-adaptive multiscale edge detector tuned to vessels of different radii. The balanced gate computes a foreground-to-background ratio $T$; images with $T < \\mu$ (here $\\mu = 0.15$) enter one queue, images with $T > \\mu$ enter another, and when both queues reach capacity $C$ the $2C$ images are concatenated along the channel dimension, forcing each mini-batch to mix foreground-poor and foreground-rich examples. The Bayesian treatment regards the LoG kernel as the prior and the module output as the posterior, maximizes the evidence lower bound (ELBO), and uses a total loss of Dice minus ELBO.","core_discovery":"On the paper's own terms, the central discovery is that a two-stream network outperforms every compared method on both aorta regions, with the largest margin on the small supra-aortic branches: SA Dice reaches 0.927 ± 0.011 against Swin-UNETR's 0.896 ± 0.019, and MA Dice reaches 0.937 ± 0.006. The gain is attributed to the LoG stream, whose five convolutional layers (kernel sizes 3, 5, 7, 9, 11) are initialized as 3D Laplacian-of-Gaussian kernels at $\\sigma$ values 0.5 to 2.5 and then optimized, so the network enhances vessels of widely different radii without manual scale selection. Bayesian parameterization of that stream gives confidence intervals that widen where boundaries are blurred. The balanced gate is reported as a major contributor: removing it drops SA Dice from 0.927 to 0.896, exactly the level of the strongest attention baseline.","pith_inferences":["The same two-stream recipe—a coarse segmentation encoder plus a trainable multiscale edge-filter stream—should transfer to other branching vessel trees such as coronary, cerebral, and pulmonary vessels, where the root-to-branch diameter ratio is similarly large.","A per-image foreground-fraction reweighting, instead of a queue that must fill before any batch is formed, might recover most of the balanced gate's effect while remaining well-defined at inference time.","The gate's queue capacity is never stated and its inference-time operation is not described, so the reported 3-point contribution of the gate cannot yet be reproduced from the paper alone; a concrete gate specification is the natural next step.","The widened confidence intervals on blurred boundaries suggest the model could serve as a pre-CFD reliability screen, flagging regions where a single reconstructed aortic surface would be misleading."],"forward_implications":["Supra-aortic branch segmentation improves by roughly 3 Dice points over the strongest attention baseline, with the largest visual gains on small-radius vessels that other methods miss.","The Bayesian LoG stream turns segmentation into a distribution: ten forward passes yield confidence intervals, and intervals widen on blurred boundaries, giving analysts a spatial map of where the prediction is reliable.","Removing the balanced gate drops SA Dice to 0.896, the same level as Swin-UNETR, indicating that foreground-background rebalancing contributes most of the branch-class gain.","CFD simulations on nine sampled segmentations show measurable flow variations, so the uncertainty intervals propagate into hemodynamic quantities such as velocity, pressure, and wall shear stress.","The added computational cost is modest at 48.88 GFLOPS, between UNETR's 41.19 and far below nnU-Net's 412.65, so the accuracy gain is not bought with an impractical model."],"supporting_citations":[{"why":"Supplies the 3D U-Net architecture used as the regular stream for overall aorta segmentation.","marker":"[2]"},{"why":"Swin-UNETR is the strongest attention baseline and the direct comparison for the reported SA Dice gain.","marker":"[15]"},{"why":"nnU-Net is the strongest CNN baseline and the reference for the computational-cost comparison.","marker":"[26]"},{"why":"Provides the Bayesian neural-network parameterization and ELBO objective used for the LoG stream.","marker":"[22]"},{"why":"The gated dual-stream design motivates the balanced gate in the LoG stream.","marker":"[17]"},{"why":"Supplies the first set of aorta training and testing volumes used in the experiments.","marker":"[23]"},{"why":"Supplies the multicenter aortic vessel tree CTA volumes with ground-truth segmentation masks.","marker":"[24]"},{"why":"Used for mesh generation in the downstream CFD demonstration.","marker":"[29]"},{"why":"Used to solve the Navier-Stokes equations in the downstream CFD demonstration.","marker":"[30]"}],"fun_headline_variants":["Bayesian LoG stream lifts aorta branch Dice by ~3%","Aorta segmentation gets Bayesian edge filter for 3% Dice gain","Hierarchical LoG Bayesian network boosts supra-aortic Dice to 0.927","Trainable LoG kernels sharpen aorta branch segmentation","Confidence intervals for aorta segmentation via Bayesian LoG"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the balanced gate can be implemented as described—two queues of unspecified capacity $C$ that concatenate $2C$ images once full—and also works at inference time on a single volume, because the paper attributes a roughly 3-point SA Dice gain to this gate yet gives no inference-time procedure.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian LoG stream lifts aorta branch Dice by ~3%","Aorta segmentation gets Bayesian edge filter for 3% Dice gain","Hierarchical LoG Bayesian network boosts supra-aortic Dice to 0.927","Trainable LoG kernels sharpen aorta branch segmentation","Confidence intervals for aorta segmentation via Bayesian LoG"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000466,"raw_usage":{"total_tokens":2336,"prompt_tokens":969,"completion_tokens":1367,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":1280}},"tokens_in":585,"tokens_out":1367,"duration_ms":10259,"temperature":1.0,"reasoning_tokens":1280,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:01:20.878739+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Force the balanced gate to open on every image by setting its queue capacity to one, retrain on the same 34 volumes, and measure SA Dice on the same 16 test volumes; if the score does not fall from 0.927 toward the 0.896 gate-off level, the gate is not carrying the claimed gain.","supporting_citations":[{"cited_title":"Semi-automatic methods in- volve manual lumen boundary and centerline annotation, achieving good performance but are labor-intensive","cited_arxiv_id":null,"evidence_quote":"Supplies the 3D U-Net architecture used as the regular stream for overall aorta segmentation."},{"cited_title":"Blood vessel segmentation algorithms - review of methods, datasets and evaluation metrics,","cited_arxiv_id":null,"evidence_quote":"nnU-Net is the strongest CNN baseline and the reference for the computational-cost comparison."},{"cited_title":"Swin UNETR: Swin transformers for semantic segmentation of brain tumors in MRI images,","cited_arxiv_id":null,"evidence_quote":"Provides the Bayesian neural-network parameterization and ELBO objective used for the LoG stream."},{"cited_title":"Pyramid scene parsing network,","cited_arxiv_id":null,"evidence_quote":"The gated dual-stream design motivates the balanced gate in the LoG stream."},{"cited_title":"ConvFormer: Com- bining CNN and transformer for medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the first set of aorta training and testing volumes used in the experiments."},{"cited_title":"Gated- SCNN: Gated shape CNNs for semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the multicenter aortic vessel tree CTA volumes with ground-truth segmentation masks."},{"cited_title":"Explanation and use of uncertainty quantified by Bayesian neural network clas- sifiers for breast histopathology images,","cited_arxiv_id":null,"evidence_quote":"Used for mesh generation in the downstream CFD demonstration."},{"cited_title":"The vascular model repository: A public resource of medical imaging data and blood flow simulation results,","cited_arxiv_id":null,"evidence_quote":"Used to solve the Navier-Stokes equations in the downstream CFD demonstration."}],"review_version":1}