{"id":"f18c7ce2-1e9b-4407-b57a-c76ebac65a95","arxiv_id":"2502.00314","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"ViLU-Net, a U-Net built with Vision LSTM blocks, reports the best segmentation scores on a new retroperitoneal tumor CT dataset and on FLARE22, but the comparison lacks error bars and omits the closest prior architecture.","lead":"This paper introduces a new CT dataset of retroperitoneal tumors and benchmarks four U-Net style segmentation networks, including a new xLSTM-based model called ViLU-Net. The authors report the best Dice scores for ViLU-Net, but without error bars, a comparison to the closest prior model, or a public dataset link.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported superiority of ViLU-Net rests on a single unvalidated split with no error bars or significance tests, and the near-identical xLSTM-UNet baseline is omitted; the gap in Table 2 may not be robust.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: the performance gap could be an artifact of the evaluation protocol, including the missing xLSTM-UNet comparison. My independent reading of the manuscript confirms this. The paper's central contribution—a new dataset plus a claim that ViLU-Net outperforms SOTA methods—depends on Table 2, but that table lacks error bars, cross-validation, split details, and statistical tests. The omission of xLSTM-UNet is especially problematic because the method section explicitly describes ViLU-Net as following xLSTM-UNet's architecture. Without that baseline, the reader cannot tell whether Vi-blocks add anything. The efficiency claim is also unquantified, but the evaluation-protocol issue is more load-bearing because it affects the primary accuracy claim. I therefore recommend keeping the reader's CONDITIONAL verdict: the paper should be accepted only if the authors add repeated split evaluation, statistical tests, and the xLSTM-UNet baseline. My concern does not change the verdict because it is the same concern the reader already raised.","tokens_in":9582,"tokens_out":2207,"duration_ms":22858,"concrete_test":"Run 5-fold cross-validation (or at least three random 80/20 splits) on the 77-case retroperitoneal dataset with identical hyperparameters for all methods, including xLSTM-UNet, and report mean ± std DSC/NSD/HD/IoU plus a paired test (e.g., Wilcoxon signed-rank) for ViLU-Net against each baseline. If the ViLU-Net advantage over nnU-Net and xLSTM-UNet does not persist across folds or is not statistically significant, the Table 2 superiority claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 2 reports one DSC/NSD/HD/IoU value per method on the retroperitoneal dataset, but §3.2.2 gives no train/validation split, no cross-validation, no variance, and no statistical test. With 77 cases (§3.2.1), a single split can easily produce a 0.03 DSC gap (0.9309 vs 0.9013 for nnU-Net) by chance, so the central claim that ViLU-Net 'demonstrates superior performance' is not yet supported. Additionally, §2 states ViLU-Net is built 'akin to xLSTM-UNet' using the same ViL blocks, yet xLSTM-UNet is never evaluated. Because ViLU-Net differs from xLSTM-UNet mainly in block arrangement, excluding this baseline means the comparison cannot isolate the proposed modification's contribution. Finally, the paper claims 'a significant reduction in complexity' but provides no parameter, FLOP, or runtime numbers anywhere. The load-bearing assumption is that the single reported evaluation is representative and complete; §3.2.2 provides no evidence for that.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper evaluates U-Net variants built from CNN, Vision Transformer, Mamba, and xLSTM components on two CT datasets: a public abdomen multi-organ dataset (FLARE 2022) and a newly introduced in-house retroperitoneal tumor dataset of 82 cases (with 77 used after preprocessing). The authors propose ViLU-Net, a U-shaped architecture that inserts Vision-LSTM (ViL) blocks into the encoder-decoder path, and report that it achieves the highest average DSC, NSD, and IoU and the lowest average HD on both datasets, while also claiming a significant reduction in complexity. The code is publicly released, and the dataset is presented as a contribution.","tokens_in":9851,"tokens_out":2855,"duration_ms":26412,"significance":"If the reported results are robust, the paper would provide two useful contributions: a new expert-annotated retroperitoneal tumor CT dataset, and evidence that an xLSTM/ViL-based U-Net variant can match or outperform transformer- and Mamba-based counterparts at lower computational cost. The study also benchmarks several recent architectures on the same protocol, which is informative for practitioners. The public code is a concrete strength. However, the headline superiority claim is supported only by single-split metric averages with no variance or significance testing, and the efficiency claim is not quantified, so the significance at this stage is mainly as a dataset release and preliminary benchmark rather than a rigorously established architectural finding.","major_comments":[{"comment":"The central claim that ViLU-Net 'demonstrates superior performance' on the retroperitoneal dataset rests on a single train/validation split with one DSC/NSD/HD/IoU value per method and no error bars, confidence intervals, cross-validation, or statistical significance tests. With only 77 cases, the 0.03 DSC gap over nnU-Net (0.9309 vs 0.9013) could easily arise from the choice of split. The authors should report k-fold cross-validation or at least bootstrapped confidence intervals and a paired significance test (e.g., Wilcoxon signed-rank or paired bootstrap) across cases to demonstrate that the improvement is not noise.","section":"§3.2.2, Table 2"},{"comment":"ViLU-Net is described as 'akin to xLSTM-UNet' and built from the same ViL blocks, yet xLSTM-UNet (reference [30], the direct architectural predecessor) is never evaluated on either dataset. Because ViLU-Net differs from xLSTM-UNet mainly in block arrangement and integration details, omitting this baseline means the comparison cannot isolate the contribution of the proposed modification. The xLSTM-UNet baseline should be included under the same training and evaluation protocol.","section":"§2 and §3.2.2"},{"comment":"The paper claims 'a significant reduction in complexity' and 'reduced complexity' (Abstract, Introduction, Conclusion) but provides no quantitative measure of efficiency anywhere: no parameter counts, FLOPs, memory usage, training time, or inference latency are reported for ViLU-Net or any baseline. This claim is currently unsupported and should be substantiated with concrete complexity measurements for all compared methods.","section":"§1 and §5"},{"comment":"The abdomen CT results contain a mathematically impossible relationship between DSC and IoU: for the same binary segmentation, IoU is always less than or equal to DSC, yet Table 1 reports DSC 0.8469 and IoU 0.9662 for nnU-Net (similarly for other rows). This indicates that the IoU values are either computed on different classes/sets (e.g., averaged over all classes rather than foreground only) or are otherwise inconsistent with the DSC computation. The metric definitions and averaging procedure must be clarified and corrected, because this casts doubt on the correctness of all reported metrics in both tables.","section":"§3.1, Table 1"}],"minor_comments":[{"comment":"The number of cases is inconsistent: the Abstract and §3.2 state the dataset has 82 cases, but §3.2.1 says 85 initial cases with 8 excluded, yielding 77 used cases. This arithmetic (85-8=77) does not match 82; please clarify whether the dataset comprises 82 or 77 cases, or state the number of included and excluded cases precisely.","section":"Abstract and §3.2.1"},{"comment":"The ViLU-Net architecture hyperparameters are not specified: the number of ViL blocks at each resolution, channel widths, patch size, and downsampling factors are absent. Figure 1 shows a schematic, but a table or explicit configuration is needed for reproducibility, especially since the implementation is based on nnU-Net's framework which may set some of these automatically.","section":"§2.2"},{"comment":"There are several typos and duplicated words, e.g., 'andreliantonextensive' (missing spaces) and 'the the Mamba-based method' in §3. A careful proofreading pass is needed.","section":"Introduction, §1"},{"comment":"The qualitative claims in the text (e.g., 'more robust and smoother segmentations', 'minimal false positives') refer to Figure 3, but the figure is not referenced with panel-specific descriptions that quantify what is being seen; adding segmentation overlays with arrows and a discussion of failure cases would strengthen the presentation.","section":"§3.2.2"}],"recommendation":"major_revision","confidential_remarks":"The core idea is reasonable and the dataset/code release are valuable, but the evidence for the superiority claim is currently too weak: a single split without error bars, no xLSTM-UNet baseline, unsupported complexity claims, and a DSC/IoU inconsistency in Table 1 that suggests a metrics bug. These are fixable with additional experiments and corrections, so I am not recommending rejection. The authors should also double-check whether the IoU averaging matches standard multi-organ evaluation practice, as this affects both tables."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — the useful thing here is the dataset, not the architecture. You get 77 annotated retroperitoneal tumor CTs with five disease categories and open-source code; that's a genuine contribution for a rare site. The benchmark across CNN, Transformer, Mamba, and xLSTM families is well motivated. But the evidence for the headline claim is shaky, and there's a reporting inconsistency that should have been caught.\n\nIn Table 1, the abdomen CT results, DSC and IoU violate the identity IoU = DSC/(2 − DSC). nnU-Net is reported with DSC 0.8469 and IoU 0.9662; those cannot both be true for binary segmentation. ViLU-Net has 0.8594 and 0.9716, same problem. Either the IoU column is not IoU or the DSC column is not Dice. That undermines the abdomen results entirely. Table 2's retroperitoneal numbers are internally consistent, so the issue may be isolated to one table, but it needs a careful re-check.\n\nOn the retroperitoneal benchmark itself: one split, no error bars, no cross-validation, no significance test. With 77 cases, the DSC gap between ViLU-Net (0.9309) and nnU-Net (0.9013) is plausible but not demonstrated to be real. The paper doesn't even state the train/validation split. Also, ViLU-Net is described as 'akin to xLSTM-UNet' and built from the same ViL blocks, but xLSTM-UNet is never included in the comparison. Without that baseline, the contribution of the rearrangement cannot be isolated. The claim of 'significant reduction in complexity' has no parameter counts, FLOPs, or runtime anywhere. Minor point: the abstract says 82 cases, the methods say 77 after exclusions; that's fine, but it should be explicit from the start.\n\nThe dataset and code are good, but the evaluation needs to be redone before any superiority claim is taken seriously. I'd still send this to review — the dataset justifies a referee's time, and the problems are fixable, not fatal to the dataset's value. The referee should push for multiple splits, error bars, the xLSTM-UNet baseline, and a corrected Table 1.","headline":"The new retroperitoneal tumor dataset is the real contribution, but the reported metrics in Table 1 are internally inconsistent and the ViLU-Net superiority claim rests on a single unvalidated split without error bars or the key xLSTM-UNet baseline.","tokens_in":10332,"tokens_out":3266,"would_cite":false,"duration_ms":31857,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A U-Net built on xLSTM blocks outperforms convolutional, transformer, and Mamba variants on retroperitoneal tumor CT segmentation, and does so with less complexity.","keywords":["Retroperitoneal tumor","CT image segmentation","U-Net","xLSTM","ViLU-Net","Mamba","Vision Transformer","Deep learning"],"falsifier":"Run a 5-fold cross-validation of ViLU-Net, nnU-Net, SwinUNETR, U-Mamba, and xLSTM-UNet on the same 77-case retroperitoneal dataset, with identical preprocessing, the same 300-epoch training schedule, and per-model hyperparameter search for each fold. If ViLU-Net's mean DSC across folds is not higher than nnU-Net's by more than the standard deviation of the fold-wise scores, or if xLSTM-UNet matches it, the claimed superiority is not established.","tokens_in":9442,"feed_emoji":"🩻","tokens_out":7152,"duration_ms":61291,"temperature":0.7,"pith_summary":"This paper argues that swapping the core of a U-Net from convolutional, transformer, or Mamba blocks to Vision xLSTM blocks yields both more accurate and cheaper segmentation of retroperitoneal tumors in CT scans. To make the case, the authors introduce a new expert-annotated CT dataset of retroperitoneal tumors (82 cases, 77 after preprocessing) and compare four U-Net-style architectures on it plus the public FLARE abdominal organ dataset. Their proposed model, ViLU-Net, reports the best average Dice Similarity Coefficient (0.9309), Normalized Surface Distance (0.9292), and Intersection over Union (0.8720) and the lowest Hausdorff distance (11.19) on the tumor dataset, ahead of nnU-Net, SwinUNETR, and U-Mamba. The authors further claim the model offers a significant reduction in complexity, which matters because transformer-based segmentation is often too memory-hungry for clinical hardware. If the claim holds, xLSTM blocks would be a practical drop-in upgrade for U-shaped segmentation networks in resource-constrained medical settings.","feed_headline":"xLSTM U-Net beats CNN, Transformer, Mamba on tumor CTs","feed_subtitle":"New model posts the top Dice score on retroperitoneal tumor segmentation at lower computational cost.","key_machinery":"The load-bearing component is the ViL (Vision xLSTM) block, built from mLSTM cells whose recurrence is $C_t = f_t C_{t-1} + i_t v_t k_t^\\top$, $n_t = f_t n_{t-1} + i_t k_t$, and $\\tilde h_t = C_t q_t / \\max(|n_t^\\top q_t|, 1)$, with hidden state $h_t = o_t \\odot \\tilde h_t$. Odd-numbered blocks scan patch tokens from top-left to bottom-right and even-numbered blocks scan in reverse, so the network sees long-range context in both directions with linear complexity in sequence length. In ViLU-Net, a convolutional stem with instance normalization and LeakyReLU produces the base feature map, two ViL blocks form each encoder stage, down-samplers halve spatial resolution while doubling channels, and up-samplers plus skip connections restore resolution before a softmax segmentation head. The mLSTM cells are what make the architecture simultaneously parallelizable, long-range, and cheap—the property the paper argues gives xLSTM an edge over ViT and Mamba companions inside U-Net.","core_discovery":"On the paper's own terms, the central discovery is that a U-shaped segmentation network whose encoder-decoder core is built from Vision xLSTM blocks—called ViLU-Net—dominates three contemporary alternatives on both a new retroperitoneal tumor CT dataset and the FLARE abdominal organ dataset. Against nnU-Net, SwinUNETR, and U-Mamba, ViLU-Net achieves the highest average DSC (0.9309 vs. 0.9013, 0.8310, and 0.7694) and lowest HD (11.19 vs. 24.64, 51.28, and 64.35) on the tumor dataset, with corresponding gains in NSD and IoU. The authors interpret this as evidence that xLSTM's exponential gating and matrix-memory recurrence capture long-range dependencies more effectively than convolutions, self-attention, or selective state spaces, while keeping linear memory cost. They also claim the model is simpler than the alternatives, positioning it as a candidate for deployment where computational resources are limited.","pith_inferences":["If the reported gains survive proper statistical evaluation, xLSTM blocks could become a standard plug-in for U-Net encoders wherever memory is the bottleneck, not just for retroperitoneal tumors.","The omission of xLSTM-UNet as a baseline means the paper has not isolated whether ViLU-Net's advantage comes from its specific modifications or from the xLSTM blocks already present in xLSTM-UNet; a head-to-head comparison would separate the two.","Since the ViL block is modality-agnostic, the same architecture could be transferred to MRI or PET tumor segmentation; a natural test would be applying ViLU-Net to a public brain-tumor or prostate MRI benchmark.","The paper's claim of reduced complexity could be quantified directly by reporting parameter counts, FLOPs, and memory per case for each baseline; that number is not in the paper and is the key missing evidence for the efficiency claim."],"forward_implications":["The xLSTM building block can replace transformer and Mamba blocks in U-shaped medical segmentation networks without losing accuracy, since ViLU-Net beats both SwinUNETR and U-Mamba on the tumor dataset.","Retroperitoneal tumor segmentation, made hard by irregular shapes and proximity to vital structures, can be automated at a level the paper argues is clinically usable, with Hausdorff distance more than halved relative to the best CNN baseline.","The released expert-annotated dataset of 82 retroperitoneal tumor cases gives the community a benchmark for a class of rare tumors where public data are scarce.","Because ViLU-Net also leads on the FLARE multi-organ abdomen dataset, the authors' claim is not limited to one tumor type; the same architecture generalizes across segmentation targets.","Lower computational complexity than transformer-based alternatives would make accurate segmentation feasible on the limited GPU hardware typical of clinical settings, a claim the paper explicitly makes."],"supporting_citations":[{"why":"Supplies the xLSTM model with exponential gating and matrix memory that defines the mLSTM cells used in ViL blocks.","marker":"[28]"},{"why":"Introduces the ViL (Vision LSTM) backbone with alternating scan directions that ViLU-Net integrates into its encoder-decoder.","marker":"[29]"},{"why":"Provides the xLSTM-UNet architecture that ViLU-Net is built from and the codebase the implementation is based on.","marker":"[30]"},{"why":"The Mamba-based baseline and another foundation for the implementation, which ViLU-Net must outperform.","marker":"[27]"},{"why":"The Swin transformer-based baseline that ViLU-Net surpasses on both datasets.","marker":"[32]"},{"why":"The nnU-Net framework used to train all models and the CNN baseline that provides the strongest competition on the tumor dataset.","marker":"[33]"},{"why":"Supplies the public FLARE22 abdominal organ dataset used as the second evaluation benchmark.","marker":"[34]"}],"fun_headline_variants":["ViLU-Net tops tumor segmentation with xLSTM core","xLSTM powers U-Net to best Dice on retroperitoneal tumors","Efficient xLSTM U-Net outshines CNN, ViT, and Mamba","ViLU-Net: xLSTM beats all in tumor CT segmentation","xLSTM U-Net wins on tumor CTs with less compute"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire superiority claim rests on a single fixed train/validation split of 77 scans, with no cross-validation, no error bars, no statistical test, and no direct comparison against xLSTM-UNet, the near-identical architecture the model is built from; if hyperparameters were not equally tuned for each baseline or the split is unrepresentative, the ranking in Table 2 could change.","fun_headline_variants_meta":{"raw":{"variants":["ViLU-Net tops tumor segmentation with xLSTM core","xLSTM powers U-Net to best Dice on retroperitoneal tumors","Efficient xLSTM U-Net outshines CNN, ViT, and Mamba","ViLU-Net: xLSTM beats all in tumor CT segmentation","xLSTM U-Net wins on tumor CTs with less compute"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000285,"raw_usage":{"total_tokens":1674,"prompt_tokens":933,"completion_tokens":741,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":641}},"tokens_in":549,"tokens_out":741,"duration_ms":6773,"temperature":1.0,"reasoning_tokens":641,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T19:26:46.468660+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a 5-fold cross-validation of ViLU-Net, nnU-Net, SwinUNETR, U-Mamba, and xLSTM-UNet on the same 77-case retroperitoneal dataset, with identical preprocessing, the same 300-epoch training schedule, and per-model hyperparameter search for each fold. If ViLU-Net's mean DSC across folds is not higher than nnU-Net's by more than the standard deviation of the fold-wise scores, or if xLSTM-UNet matches it, the claimed superiority is not established.","supporting_citations":[{"cited_title":"xLSTM-UNet can be an Effective 2D & 3D Medical Image Segmentation Backbone with Vision-LSTM (ViL) better than its Mamba Counterpart","cited_arxiv_id":"2407.01530","evidence_quote":"Provides the xLSTM-UNet architecture that ViLU-Net is built from and the codebase the implementation is based on."},{"cited_title":"Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images,","cited_arxiv_id":null,"evidence_quote":"The Swin transformer-based baseline that ViLU-Net surpasses on both datasets."},{"cited_title":"nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"The nnU-Net framework used to train all models and the CNN baseline that provides the strongest competition on the tumor dataset."}],"review_version":1}