{"id":"a135d324-c288-4484-9ea4-aeab0bdb5582","arxiv_id":"1908.01373","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An ACWE-inspired unsupervised neural network with morphological pooling layers segments 3D microvasculature and outperforms supervised baselines in reported experiments.","lead":"A new unsupervised deep learning method segments 3D blood vessel images by using a loss inspired by active contours and adding novel morphological pooling layers. The authors report it beats supervised vessel segmentation networks on both same-dataset and cross-dataset tests, with code available.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported margins over DeepVess are tiny and no error bars are given; with a single test volume and no multiple seeds, the superiority claim is not statistically established.","rationale":"I focus on evaluation reliability because the paper's own strongest claim is explicitly comparative: the method achieves the best AP, F1, JI, DICE, and mIoU on the two datasets and in cross-dataset transfer. For that claim to hold, the reported margins must be distinguishable from experimental noise. The DeepVess test set is a single annotated volume and the VesselNN test set is only two volumes; the paper reports no variance, no number of runs, and no significance test. On DeepVess, the margins over the strongest baseline are as small as 0.004 in JI and 0.009 in F1, making the superiority claim fragile even before considering that the baselines are reimplementations. The reader's named weakest assumption, the bright-vessel prior encoded in Lrank, is genuine but less load-bearing for the central claim: all datasets used in the paper are two-photon images with bright contrast agents, so the assumption matches the evaluated scope. It limits generality but does not directly undermine the in-domain comparison. The evaluation-reliability concern does undermine the comparison itself. I therefore keep the verdict at CONDITIONAL, unchanged from the reader, but for a different primary reason. A concrete statistical reanalysis, as described in the test, would settle whether the claim survives.","tokens_in":12755,"tokens_out":9231,"duration_ms":94106,"concrete_test":"Re-run Tables 1-3 with at least 5 random initializations for every method and report mean plus 95% CI; for VesselNN report per-test-volume scores (only 2 volumes) and for DeepVess use block bootstrap over z-slices or subvolumes of the single annotated volume. If the 95% CI for Ours minus DeepVess on AP, F1, JI, DICE, or mIoU contains 0, the superiority claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is that the ACWE-loss network achieves the best AP, F1, JI, DICE, and mIoU on DeepVess and VesselNN and in cross-dataset transfer. For that claim to be secure, the observed margins must be distinguishable from evaluation noise. They are not: DeepVess is a single 200x256x256 volume split into train/test (Sec. 5.1), and VesselNN has only 2 test volumes. On DeepVess the reported gaps over DeepVess are AP 0.909 vs 0.889, F1 0.829 vs 0.820, JI 0.811 vs 0.807, DICE 0.829 vs 0.820, mIoU 0.838 vs 0.828. No error bars, number of runs, or significance tests are reported anywhere in Tables 1-3. Under run-to-run and train/test-split variability, such differences may be within the noise level. The FT protocol (Sec. 4.4) further gives the unsupervised method test-time adaptation and extra compute; while Ours-without-FT also often leads in Table 3, the fairness of the headline comparison is not fully addressed. The bright-vessel assumption (Eq. 22, Lrank) is a real limitation, but it is explicitly scoped by the paper's datasets; the lack of statistical support threatens the stated outperformance claim even within that scope.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents an unsupervised deep learning method for 3D blood vessel segmentation. The key idea is to replace the level-set evolution of morphological Active Contours Without Edges with a differentiable loss and with morphological pooling layers embedded in a 3D ResNet-based autoencoder. The training objective combines an ACWE-inspired image attachment loss, a ranking loss that enforces brighter vessels, a reconstruction loss, a minimal-segmentation loss, and two disjunctive losses. Experiments on the DeepVess and VesselNN datasets, including cross-dataset transfer, compare against supervised baselines (VesselNN, DeepVess) and classical methods (VIDA, Morph-ACWE); further qualitative results are presented on a 4D intravital dataset. The paper claims state-of-the-art unsupervised performance, and the ablation study attributes the gain to the proposed ACWE loss and the ranking loss.","tokens_in":13107,"tokens_out":6692,"duration_ms":63996,"significance":"The idea of designing a network architecture and loss that mimic morphological ACWE is novel and well motivated, and the ablation study (Table 4) demonstrates that the ACWE loss and the ranking loss are important for performance. The method addresses a real need: unsupervised segmentation across imaging-domain shifts, and the transductive fine-tuning protocol (Sec. 4.4) is an interesting way to exploit unlabeled test data. However, the quantitative evidence for the headline claim is not yet statistically convincing, and an internal inconsistency in the reported results must be resolved before the strengths can be fully credited.","major_comments":[{"comment":"Table 4 reports JI=0.708 for the full method \"OursL\", while Table 1 reports JI=0.811 for the same method on the same dataset; the other overlapping metrics (AP, F1, DICE, mIoU) match exactly. This is an internal contradiction. Please correct the value or explain the discrepancy (e.g., a different evaluation protocol or train/test split). Without this, the ablation table cannot be considered reliable.","section":"Sec. 5.2, Table 4 vs. Table 1"},{"comment":"The central claim that the method outperforms supervised baselines is based on small margins without any measure of uncertainty. On DeepVess, which has a single test volume, the gains over DeepVess are AP 0.909 vs. 0.889, F1 0.829 vs. 0.820, JI 0.811 vs. 0.807. No error bars, number of runs, or significance tests are provided anywhere in the paper. Please report mean ± std over multiple training runs (e.g., different random seeds) and, if possible, bootstrap confidence intervals or paired significance tests over the test volumes. With the current evidence, the observed differences are within plausible run-to-run and train/test-split variability, so the stated outperformance claim is not established.","section":"Sec. 5.2, Tables 1-3"},{"comment":"The FT (unsupervised fine-tuning) protocol uses the test data, giving the method an information and compute advantage over the supervised baselines, which are not fine-tuned. While the paper marks FT explicitly, the narrative in the abstract and in Sec. 5.2 (\"our unsupervised method is able to outperform such previous methods\") relies on the FT results. Please either present the no-FT results as the primary cross-domain comparison or provide a fair baseline that also uses test data (e.g., test-time adaptation or a transductive variant). The no-FT rows in Table 3 are often favorable, but the current presentation makes the headline comparison difficult to interpret fairly.","section":"Sec. 4.4, Table 3"},{"comment":"The paper states that the reimplementations of DeepVess and VesselNN achieve the same level of results as the original implementations, but it does not report the original numbers or the details of the reimplementation procedure (e.g., hyperparameters, training duration, preprocessing beyond normalization). Without this information, the reader cannot verify that the baselines are not disadvantaged. Please include a side-by-side comparison with the original published metrics and provide the reimplementation settings in the supplementary material or appendix.","section":"Sec. 5 (baseline reimplementations)"}],"minor_comments":[{"comment":"The first affiliation contains a typo: \"School of Computer Sceince\" should be \"School of Computer Science\".","section":"Author affiliations"},{"comment":"Equation (12) is garbled: the notation \"SI(IS(... (SI(IS| {z } SI◦IS µtimes\" is not readable and should be typeset properly, e.g., as an iteration of the composite operator (SI ◦ IS) applied µ times.","section":"Eq. (12)"},{"comment":"The sentence \"the F1 score and the Average Precision are the more informative metrics\" should likely read \"the most informative metrics\".","section":"Sec. 5.2"},{"comment":"The qualitative evaluation on the 4D-NVIV dataset is only descriptive; the expert-preference statement lacks a protocol, the number of experts, and any measure of agreement. A more formal evaluation would strengthen the subjective claim.","section":"Sec. 5.2, 4D-NVIV evaluation"},{"comment":"For reproducibility, the paper should state the batch size, the number of training epochs or iterations, and the learning-rate schedule in addition to the fixed loss weights and optimizer.","section":"Sec. 4.3"},{"comment":"It would be helpful to state explicitly that the default loss in Eq. (21) uses the smoothed output S and that gradients therefore flow through the morphological pooling layers; the ablation that replaces S with S̄ is clear, but the default choice is currently implicit.","section":"Eq. (21)"}],"recommendation":"major_revision","confidential_remarks":"The paper has a strong conceptual contribution and the ablation supports the role of the proposed losses. However, the lack of statistical evidence for the headline outperformance claim and the internal inconsistency between Table 1 and Table 4 are load-bearing issues. These are fixable with additional experiments and careful reporting, so I recommend major revision rather than rejection. The authors should also be asked to clarify the exact evaluation protocol of the ablation study and to ensure the baseline comparisons are transparent."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look. The core idea is new and sensible: instead of using active contours as a post-processing or supervised auxiliary loss, the authors turn the ACWE energy update into an unsupervised loss and implement the morphological curvature operator as fixed, differentiable pooling layers. That integration is not in the prior work they cite, which all used supervised losses. The ablations are the strongest part of the paper — they show L_AC alone does most of the work, L_rank matters, and replacing the smoothed output S with the raw S-hat in the loss hurts clearly. That is evidence the design is doing what it claims, not just a stack of tricks. Shipping code and PyTorch reimplementations of both baselines is also a real plus and makes the comparison reproducible.\n\nThe soft spot is the evaluation. DeepVess is a single 200x256x256 volume split into train/test; VesselNN has only two test volumes. On DeepVess the reported margins over their reimplemented DeepVess are AP 0.909 vs 0.889, F1 0.829 vs 0.820, JI 0.811 vs 0.807, DICE 0.829 vs 0.820, mIoU 0.838 vs 0.828. With no error bars, no number of runs, and no significance test, those deltas could easily be split or run-to-run noise. The cross-dataset gaps are larger and more convincing, and the non-FT versions also lead there, so the FT fairness concern is secondary — but the FT protocol itself is a real asymmetry that should at least be discussed more carefully. Their claim in the text of \"significantly outperforms\" is not supported by the reported statistics; at best it is supported in the cross-dataset direction on a handful of volumes, and only descriptively.\n\nThe bright-vessel assumption in L_rank is explicitly tied to the datasets, which have a fluorescent contrast agent. The paper scopes itself to that, so I would not call it a flaw, but it is worth flagging for any reader who generalizes. The hand-set weights are also a minor concern — fixed across experiments, but they are still free parameters and there is no sensitivity analysis.\n\nNet: the method is interesting, the mechanism is well motivated, and the ablations show the machinery works. The evaluation is too thin for the \"outperforms\" headline. I would recommend peer review — this deserves a serious referee — but the revision needs error bars, multiple seeds, and ideally a test set larger than one or two volumes. For someone working on unsupervised medical image segmentation, this is a paper to read and probably to cite, with the evaluation caveats in mind.","headline":"A genuinely new unsupervised vessel-segmentation loss plus morphological pooling layers, but the headline superiority claim rests on very small test sets and no error bars.","tokens_in":13540,"tokens_out":665,"would_cite":true,"duration_ms":8842,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An unsupervised network using an active-contours loss outperforms supervised vessel segmenters on both benchmark datasets and across datasets.","keywords":["microvascular segmentation","unsupervised segmentation","active contours without edges","morphological pooling","three-dimensional microscopy","deep learning","cross-dataset generalization","two-photon microscopy"],"falsifier":"A direct test: apply the trained network to the same volumes with intensities inverted, so vessels become darker than the background. If the central claim is right, the segmentation should invert and F1 should collapse; if it does not, the stated ranking-loss mechanism is not the actual driver.","tokens_in":12583,"feed_emoji":"🩸","tokens_out":4842,"duration_ms":51354,"temperature":0.7,"pith_summary":"This paper tries to show that blood-vessel segmentation in 3D microscopy can be done without any labeled training data, and that the resulting unsupervised network beats supervised segmenters both on their own datasets and when the imaging domain shifts. It converts the classic Active Contours Without Edges (ACWE) energy into a differentiable loss and replaces ACWE's morphological curvature operators with fixed pooling layers inside a U-shaped encoder-decoder. Because the method needs no labels, it can also be fine-tuned directly on unlabeled test volumes, which further improves cross-dataset performance. If correct, it would let vascular analysis be applied quickly to new microscopes, contrast agents, and imaging conditions that would otherwise require fresh annotation.","feed_headline":"Unsupervised network beats supervised vessel segmentation","feed_subtitle":"An active-contours loss lets an unlabeled network outscore supervised rivals and transfer across datasets.","key_machinery":"The load-bearing mechanism is the morphological ACWE pipeline split into two network ingredients. The image-attachment term $\\Gamma$ becomes the active contour loss $L_{AC}$, computed on the intermediate segmentation $\\bar S$; the ranking loss $L_{\\text{rank}}=\\exp(c_2-c_1)$ encodes the prior that vessels are bright; and the curvature operator $(SI\\circ IS)^\\mu$ is realized by morphological pooling layers that take masked max/min over nine $3\\times3\\times3$ structuring elements, applied $\\mu=3$ times to smooth $\\bar S$ into the final mask $S$. The ablation shows both $L_{AC}$ and $L_{\\text{rank}}$ are necessary, and replacing $S$ with $\\bar S$ in the loss causes tearing artifacts.","core_discovery":"The paper claims that an unsupervised network can outperform supervised approaches for microvascular segmentation. The segmentation network maps an input intensity volume $I$ to a soft mask $S$, and is trained by minimizing an ACWE-inspired loss in which each voxel is penalized when it disagrees with the sign of the image-attachment term $\\Gamma = \\|\\nabla \\bar S\\|_1(\\alpha(I-c_1)^2-\\beta(I-c_2)^2)$, where $c_1,c_2$ are the intensity means inside and outside the current mask and, following the morphological ACWE literature, $\\alpha=1,\\beta=2$. A ranking loss $\\exp(c_2-c_1)$ forces the interior mean above the exterior mean, and non-learned morphological pooling layers implement the $SI\\circ IS$ curvature smoothing. On the DeepVess and VesselNN datasets the method reports the best F1, AP, JI, DICE, and mIoU among the compared methods, and in cross-dataset transfer the gap over supervised baselines grows; unsupervised fine-tuning on the test volume improves results further.","pith_inferences":["The contrast sign is hard-coded, so a natural next step would be to let the network learn whether vessels are bright or dark from the data rather than assuming brightness; the authors do not test inverted-contrast data.","Because the loss only needs intensity statistics and a smoothness prior, the same architecture could be adapted to other bright tubular structures, such as plant vasculature or engineered microfluidics, though the paper does not claim this.","The method's success in transductive fine-tuning suggests a broader recipe: unsupervised losses can serve as a domain-adaptation layer on top of any segmentation network, a use the paper mentions but does not develop."],"forward_implications":["On both benchmark datasets, the method reports the best F1, AP, JI, DICE, and mIoU among the compared supervised and unsupervised baselines.","When trained on one dataset and tested on another, the unsupervised method maintains a larger advantage over supervised baselines, and unsupervised fine-tuning on unlabeled target data (Ours-FT) yields the highest cross-dataset scores.","Without any labels, the method can be applied directly to low-SNR 4D intravital movies; expert reviewers preferred its output over DeepVess in every presented case.","The ablation indicates the ACWE loss alone already produces competitive segmentation, while adding the ranking loss and morphological smoothing gives the full gain."],"supporting_citations":[{"why":"Defines the ACWE energy functional and the inside/outside means $c_1,c_2$ that the paper's loss is built on.","marker":"[4]"},{"why":"Provides the morphological ACWE algorithm, including the curvature operator $SI\\circ IS$ and the image-attachment term $\\Gamma$ that the network layers and loss mimic.","marker":"[27]"},{"why":"DeepVess is the supervised state-of-the-art baseline and one of the benchmark datasets the method must beat.","marker":"[13]"},{"why":"VesselNN is the other supervised baseline and benchmark dataset used for within-dataset and cross-dataset comparisons.","marker":"[32]"},{"why":"The 3D U-Net result is an earlier supervised baseline listed in the DeepVess dataset comparison table.","marker":"[6]"},{"why":"VIDA provides an unsupervised baseline and was used to generate expert segmentations for the datasets.","marker":"[34]"},{"why":"ResNet34 is the encoder backbone whose intermediate blocks feed the decoders through skip connections.","marker":"[15]"},{"why":"PySight is the imaging source for the 4D intravital dataset used in the qualitative low-SNR evaluation.","marker":"[14]"}],"fun_headline_variants":["Unsupervised vessel segmentation with active contours outperforms supervised","No labels needed: unsupervised network beats supervised vessel segmentation","Active-contours loss lets unsupervised net beat supervised vessel segmentation","Unsupervised network mimics active contours to outdo supervised vessel segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes vessels appear brighter than their surroundings, because the ranking loss pushes the mean intensity inside the segmentation above the mean outside; with inverted contrast it would likely segment the background instead.","fun_headline_variants_meta":{"raw":{"variants":["Unsupervised vessel segmentation with active contours outperforms supervised","No labels needed: unsupervised network beats supervised vessel segmentation","Active-contours loss lets unsupervised net beat supervised vessel segmentation","Unsupervised network mimics active contours to outdo supervised vessel segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000588,"raw_usage":{"total_tokens":2755,"prompt_tokens":936,"completion_tokens":1819,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":1751}},"tokens_in":552,"tokens_out":1819,"duration_ms":12607,"temperature":1.0,"reasoning_tokens":1751,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:14:36.549963+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test: apply the trained network to the same volumes with intensities inverted, so vessels become darker than the background. If the central claim is right, the segmentation should invert and F1 should collapse; if it does not, the stated ranking-loss mechanism is not the actual driver.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the ACWE energy functional and the inside/outside means $c_1,c_2$ that the paper's loss is built on."},{"cited_title":"Marquez-Neila, L","cited_arxiv_id":null,"evidence_quote":"Provides the morphological ACWE algorithm, including the curvature operator $SI\\circ IS$ and the image-attachment term $\\Gamma$ that the network layers and loss mimic."},{"cited_title":"Deep convolutional neural networks for segmenting 3D in vivo multiphoton images of vasculature in Alzheimer disease mouse models","cited_arxiv_id":"1801.00880","evidence_quote":"DeepVess is the supervised state-of-the-art baseline and one of the benchmark datasets the method must beat."},{"cited_title":"C ¸ ic ¸ek, A","cited_arxiv_id":null,"evidence_quote":"The 3D U-Net result is an earlier supervised baseline listed in the DeepVess dataset comparison table."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"VIDA provides an unsupervised baseline and was used to generate expert segmentations for the datasets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ResNet34 is the encoder backbone whose intermediate blocks feed the decoders through skip connections."},{"cited_title":"Har-Gil, L","cited_arxiv_id":null,"evidence_quote":"PySight is the imaging source for the 4D intravital dataset used in the qualitative low-SNR evaluation."}],"review_version":1}