{"id":"cc0eb812-6a9b-48ef-9f79-5a7598a9b5e3","arxiv_id":"2501.07017","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"UNETVL, a 3D segmentation model replacing UNETR's transformer encoder with Vision-LSTM and Chebyshev KAN layers, achieves competitive Dice scores on ACDC and AMOS2022.","lead":"The authors combine Vision-LSTM with Chebyshev polynomial layers inside a UNETR-style encoder to create UNETVL, a 3D medical image segmentation model. They report Dice improvements over the UNETR baseline on cardiac MRI and abdominal CT benchmarks, including the top mean score on the AMOS2022 dataset.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline deltas vs UNETR are not apples-to-apples: Table 1's UNETR baselines come from the original study while UNETVL runs in nnU-Net; the AMOS SOTA gap over CoTr is 0.55 without error bars.","rationale":"The reader's weakest assumption—that comparability of baselines is the key risk—is well placed. I sharpen it by pointing to the Table 1 caption: the UNETR scores are explicitly not from the nnU-Net revisited study, while UNETVL is trained inside nnU-Net. The paper's own claim of 'uniform conditions' applies to the revisited-study comparators, not to the UNETR baseline used for the headline improvement percentages. This is load-bearing because the abstract and results section emphasize the 7.3% and 15.6% gains over UNETR. A second, related weakness is the AMOS SOTA claim: the margin over CoTr is only 0.55 Dice points and no uncertainty is reported, so 'highest mean Dice' is not statistically supported. I do not see a fatal flaw in the architecture itself; the ablation study is suggestive, and the within-pipeline comparisons to several strong baselines support the usefulness of the proposed design. However, the quantitative superiority claims need protocol-aligned baselines and error bars before the headline magnitudes can be trusted. The conditional verdict remains appropriate.","tokens_in":6475,"tokens_out":5427,"duration_ms":50543,"concrete_test":"Re-run the UNETR baseline inside the same nnU-Net pipeline used for UNETVL (same oversampling, patch size, normalization, augmentation, optimizer, epoch count, and 5-fold splits) on ACDC and AMOS post-challenge Task 2, and report per-fold Dice with mean±std. If the re-run UNETR mean is materially above 85.34/76.59, the 7.3%/15.6% improvement claims are artifacts of protocol mismatch; additionally report fold-level paired differences between UNETVL and CoTr on AMOS to see whether the 0.55-point SOTA gap survives variance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim—7.3% on ACDC and 15.6% on AMOS over UNETR—rests on a baseline that is not produced under the same protocol as UNETVL. Table 1's caption says 'All results, except for UNETR and SAM3D, are from the nnUNet revisited work.' UNETVL itself is implemented inside nnU-Net, so its preprocessing, augmentation, oversampling, and 5-fold evaluation follow the nnU-Net pipeline. The UNETR numbers (85.34 and 76.59) are instead legacy values from the original UNETR study. The paper's Section 3 statement that this setup 'ensures a fair comparison of UNETVL with other methods under uniform conditions, with the network architecture as the sole variable' is therefore not true for the UNETR comparison that headlines the abstract. If UNETR were re-run inside the same nnU-Net pipeline, the deltas could shrink or even change sign. In addition, the AMOS 'highest mean Dice' claim is 88.57 vs 88.02 for CoTr—a 0.55-point difference with no error bars, per-fold values, or significance test reported; this is within the typical spread of 5-fold cross-validation on these benchmarks. The architecture and the within-pipeline comparisons to SwinUNETR etc. are plausible, but the magnitude of the stated improvement over UNETR and the SOTA designation are not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UNETVL, a 3D medical image segmentation architecture that replaces the ViT encoder of UNETR with Vision-LSTM (ViL) blocks and injects Chebyshev KAN layers into the up- and down-projection paths. The method is implemented inside the nnU-Net framework and evaluated on the ACDC and AMOS 2022 post-challenge Task 2 datasets. The authors report mean Dice scores of 91.59 on ACDC and 88.57 on AMOS, claim improvements of 7.3% and 15.6% over UNETR, and state that the AMOS result is the highest among the compared methods. Ablation studies on the ACDC dataset are used to justify the Chebyshev KAN choice and the latent dimension.","tokens_in":6780,"tokens_out":4205,"duration_ms":39874,"significance":"If the reported results hold under uniform evaluation, the paper would present a strong and practical 3D segmentation backbone, with the code release and the nnU-Net-based implementation being concrete strengths. However, the central quantitative claims are currently not established because the UNETR baseline is taken from an external study with a different protocol, and the AMOS state-of-the-art claim rests on a 0.55-point margin without any uncertainty quantification. The architecture itself is plausible, and the within-pipeline comparisons against SwinUNETR and other modern baselines are informative, but the headline deltas and the SOTA designation need more rigorous support before the paper can be accepted.","major_comments":[{"comment":"The claim that the nnU-Net framework 'ensures a fair comparison ... with the network architecture as the sole variable' is not true for the UNETR baseline. Table 1's caption states that all results except UNETR and SAM3D come from the nnU-Net revisited study, while UNETVL is trained using nnU-Net preprocessing, augmentation, oversampling, and 5-fold evaluation. The UNETR Dice scores (85.34 on ACDC, 76.59 on AMOS) are legacy numbers from the original UNETR paper, which used a different protocol. Consequently, the 7.3% and 15.6% improvements advertised in the abstract are cross-protocol comparisons and may be artifacts of preprocessing or training differences rather than architectural superiority. The authors should rerun UNETR inside the nnU-Net pipeline, or explicitly state that the UNETR comparison is not apples-to-apples and remove the quantitative improvement claims from the abstract.","section":"Section 3, Implementation Details; Table 1"},{"comment":"The claim that UNETVL 'achieves the highest mean Dice score on the AMOS 2022 post-challenge Task 2 dataset' is not statistically supported. The margin over CoTr is only 0.55 percentage points (88.57 vs. 88.02), yet no per-fold results, standard deviations, confidence intervals, or significance tests are reported anywhere in the paper. With 5-fold cross-validation, a difference of this size can easily arise from fold-to-fold variability. The authors should report per-fold scores and some measure of uncertainty before making a SOTA claim.","section":"Section 4, Table 1"},{"comment":"The ablation table is presented without clarifying that it uses a different evaluation protocol than the final results: the caption says 'test on ACDC cross validation fold 1 after 100 epochs', whereas Table 1 reports 5-fold cross-validation results. As a result, the reader cannot directly compare the ablation numbers (e.g., 89.79 for w/ KAN, K=384 in Table 2) with the main result (91.59 in Table 1), and the conclusion that 'increasing the latent dimension results in better segmentation' is only supported on a single fold after 100 epochs. The authors should either run the full ablation over all 5 folds and the full training schedule, or explicitly state that the ablation is preliminary and only indicative.","section":"Section 4, Table 2"},{"comment":"The selection of Chebyshev polynomials as the univariate function in the KAN layer is made using 5-fold cross-validation on the ACDC dataset, and the final UNETVL model is then evaluated on the same ACDC benchmark. This creates a selection bias: the function choice is tuned to ACDC and the reported ACDC performance of the full model is not an unbiased estimate. Moreover, the difference between Chebyshev (86.12) and MLP (85.88) in Table 3 is only 0.24 percentage points, and no error bars are provided. The paper's claim that Chebyshev 'consistently demonstrated superior performance' is therefore not firmly established. The authors should report the variance of these estimates, evaluate the function choice on a held-out dataset (e.g., AMOS), or temper the claim.","section":"Section 4, Choice of Univariate Functions for KAN"},{"comment":"No error bars, standard deviations, or statistical significance tests are provided for any of the main segmentation results. Given that the paper's central claims include a 7.3% improvement on ACDC and a 0.55-point SOTA margin on AMOS, the absence of uncertainty quantification is a load-bearing omission. The authors should report per-fold Dice scores and at least the standard deviation across folds for all methods they compare, especially for their own model.","section":"Throughout; Sections 3 and 4"}],"minor_comments":[{"comment":"The tensor T is defined as having shape R^{N x K x (degree+1)}, and the summation is written over i and j, but the index j is not explicitly defined in the text as the polynomial-degree index; please add a sentence for clarity.","section":"Section 2.3, Eq. (1)"},{"comment":"The description of the mLSTM block would benefit from equations or a precise pointer to the xLSTM paper; as written, the memory-cell mechanism is described only verbally, which makes the architecture hard to reproduce from the paper alone.","section":"Section 2.2"},{"comment":"The right panel of Figure 1 is low-contrast and the labels 'Even', 'False', 'Flip', and the KAN projection annotations are difficult to read; please enlarge the figure or split it into separate panels.","section":"Figure 1"},{"comment":"There is a formatting issue: the caption appears as 'Table 3 .' and the table body is not well spaced; also, the rows for MLP, B-Spline, Gaussian RBF, and Chebyshev should be visually separated.","section":"Table 3"},{"comment":"The statement that UNETVL 'significantly mitigates over-segmentation' is not quantified; Figure 2 shows qualitative examples, but a quantitative measure such as boundary F1, surface Dice, or a per-structure over-segmentation metric would support the claim.","section":"Section 5, Discussion"},{"comment":"The paper uses both 'Kolmogorov–Arnold Networks (KAN)' and 'Kolmogorov-Arnold Network' inconsistently; please unify the terminology and check singular/plural usage.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid architectural idea and a reasonable within-pipeline comparison against several modern baselines, but the headline claims are overstated relative to the evidence. The most important issue is the protocol mismatch for the UNETR baseline: the authors should either rerun UNETR under the nnU-Net pipeline (which is feasible given their code base) or clearly reframe the comparison as cross-protocol and remove the specific percentage improvements from the abstract. In addition, adding per-fold results and standard deviations would address the SOTA claim on AMOS. If the authors make these changes, the paper could become acceptable; in its current form, the central quantitative conclusions are not established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper proposes UNETVL, a UNETR variant that swaps the ViT encoder for Vision-LSTM blocks and replaces the MLP projections with a Chebyshev KAN. That's a straightforward, reasonable combination, and the paper is honestly written around it. The work is incremental, but not every paper needs to be a breakthrough to be useful.\n\nWhat it does well: the implementation inside nnU-Net is the right way to control for preprocessing and augmentation, and the comparisons to SwinUNETR, nnFormer, and others that were also run in the nnU-Net-revisited framework are the credible part of the evaluation. The univariate-function ablation on U-KAN is a sensible sanity check, even if it's on a different architecture. Code is promised.\n\nThe soft spots are real and they matter. First, the headline deltas — 7.3% on ACDC, 15.6% on AMOS — compare against UNETR numbers taken from the original UNETR paper, not from a re-run inside nnU-Net. The claim in Section 3 that the setup 'ensures a fair comparison with the network architecture as the sole variable' is simply not true for that baseline. Second, the AMOS 'highest mean Dice' claim is 88.57 vs CoTr's 88.02, a 0.55-point gap with no error bars or significance test. That is within the usual spread of 5-fold CV on these benchmarks. Third, the main ablation table (Table 2) is run on a single fold after 100 epochs, which doesn't match the reported final results in Table 1, so it doesn't cleanly demonstrate that KAN is the cause of the gain. Fourth, the choice of Chebyshev was made on ACDC via the U-KAN experiments, and then evaluated on ACDC — a mild selection bias, not a fatal one.\n\nNone of this kills the architecture. The within-pipeline comparisons suggest ViL is a competitive encoder, and the KAN swap is a plausible tweak. But the public claims overstate what is established. The paper deserves a serious referee, provided the referee pushes for a re-run of UNETR under identical conditions, error bars on all headline numbers, and a clearer ablation protocol. If the authors fix those, this becomes a solid contribution to the incremental-architecture literature.\n\nMy take: send it to review, but expect major revision.","headline":"A sensible incremental architecture paper whose headline gains over UNETR are not apples-to-apples, and whose AMOS 'SOTA' gap is within noise; worth refereeing but needs protocol fixes.","tokens_in":7314,"tokens_out":1935,"would_cite":false,"duration_ms":19639,"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 UNETVL, a U-Net with a Vision-LSTM encoder and Chebyshev-KAN projections, raises mean Dice by 7.3% on cardiac MRI and 15.6% on abdominal CT versus UNETR, reaching the top compared mean on AMOS.","keywords":["3D medical image segmentation","Vision-LSTM","Chebyshev KAN","UNETR","U-Net encoder-decoder","ACDC cardiac MRI","AMOS 2022 abdominal CT","long-range dependency modeling"],"falsifier":"Train the UNETR baseline and UNETVL with identical data splits, preprocessing, augmentation, batch size, optimizer, and epoch counts, then compare per-fold mean Dice with standard deviations. If the reported 7.3% (ACDC) and 15.6% (AMOS) gaps shrink to roughly 2% or less, the claimed advantage is a protocol artifact rather than an architectural effect.","tokens_in":6259,"feed_emoji":"🩻","tokens_out":11617,"duration_ms":100456,"temperature":0.7,"pith_summary":"This paper tries to establish that a 3D U-Net segmentation model with a Vision-LSTM encoder and Chebyshev-KAN projection layers outperforms the ViT-based UNETR it modifies, reporting mean Dice rising from 85.34% to 91.59% on the ACDC cardiac MRI benchmark and from 76.59% to 88.57% on the AMOS 2022 post-challenge Task 2 abdominal CT benchmark. On AMOS, the model reaches the highest mean Dice among compared methods. The reason to care is that Vision-LSTM models long-range dependencies with linear rather than quadratic complexity, while Chebyshev KAN sharpens boundary detail and curbs over-segmentation. The paper also reports an ablation showing Chebyshev beats MLP, B-spline, and Gaussian RBF as the KAN univariate function in a lightweight segmentation model.","feed_headline":"Vision-LSTM encoder lifts 3D organ Dice by up to 15.6%","feed_subtitle":"It scores 91.59% Dice on cardiac MRI and tops abdominal CT at 88.57% mean Dice among compared models.","key_machinery":"The load-bearing objects are the ViL block pair and the Chebyshev KAN layer. A ViL block alternates forward and backward mLSTM layers, whose matrix-valued memory cell lets the encoder accumulate long-range spatial context over patch tokens with linear cost in sequence length. The Chebyshev KAN layer replaces the MLP in the ViL up and down projections: Chebyshev polynomials $T_0(x)=1$, $T_1(x)=x$, and $T_m(x)=2xT_{m-1}(x)-T_{m-2}(x)$ are combined with a learnable coefficient tensor $C$ of shape $K \\times O \\times (d+1)$ by Einstein summation to produce the projected feature map. Together they give the encoder long-range memory and the projection paths higher nonlinear capacity.","core_discovery":"On its own terms, the paper's discovery is that a U-Net built from Vision-LSTM encoder blocks, with Chebyshev-KAN projections in place of MLP projections, segments 3D medical volumes more accurately than the ViT-based UNETR it modifies: mean Dice rises from 85.34% to 91.59% on ACDC and from 76.59% to 88.57% on AMOS 2022 post-challenge Task 2, the highest mean among compared methods on the latter. The authors attribute this to bidirectional mLSTM memory capturing long-range spatial context with lower computational cost than self-attention, and to Chebyshev polynomials giving the projection layers stronger nonlinear approximation. They position the model as an efficiency-accuracy alternative to both convolutional and transformer encoders, with the network architecture as the only variable in an otherwise standardized pipeline.","pith_inferences":["Because several comparison-table baselines were taken from a prior study rather than rerun here, the 7.3% and 15.6% gains should be read as provisional until a same-protocol head-to-head confirms them.","The Chebyshev-versus-MLP advantage was isolated in a lightweight segmentation model on ACDC; whether it transfers to other organs, modalities, or full UNETVL scales is untested.","ViL's linear complexity suggests the encoder could be pushed to larger patch counts or higher input resolutions than the tested settings.","The qualitative over-segmentation reduction could be made quantitative with surface Dice or Hausdorff distance, which would strengthen the clinical claim."],"forward_implications":["A ViL-based encoder gives a transformer-free way to model long-range spatial dependencies in 3D volumes with linear rather than quadratic complexity.","Chebyshev KAN projections can be swapped into ViL up and down projection layers with a one-line change, raising Dice in the reported settings.","On the AMOS 2022 post-challenge Task 2 dataset, the architecture reaches the highest mean Dice among compared methods, supporting abdominal multi-organ segmentation.","The ablations show that raising the latent dimension $K$ improves accuracy at higher parameter counts, giving an explicit accuracy-cost trade-off for deployment."],"supporting_citations":[{"why":"It supplies the UNETR baseline Dice values and the encoder-decoder skeleton that UNETVL modifies.","marker":"[1]"},{"why":"It supplies the Vision-LSTM encoder blocks, built from bidirectional mLSTM layers, that replace the ViT encoder.","marker":"[4]"},{"why":"It supplies the Chebyshev polynomial KAN layer used in the up and down projection paths.","marker":"[9]"},{"why":"It supplies the comparison-table baseline scores and the rationale for choosing the two benchmark datasets.","marker":"[8]"},{"why":"It supplies the self-configuring segmentation pipeline whose preprocessing and augmentation keep training conditions uniform.","marker":"[10]"},{"why":"It introduces the Kolmogorov-Arnold Network idea that Chebyshev KAN specializes.","marker":"[5]"},{"why":"It provides the ACDC cardiac MRI benchmark and its manual anatomical labels used for evaluation.","marker":"[6]"},{"why":"It provides the AMOS 2022 post-challenge Task 2 abdominal CT benchmark and its manual labels.","marker":"[7]"}],"fun_headline_variants":["UNetVL lifts 3D organ Dice up to 15.6% over UNETR","Chebyshev-KAN Vision-LSTM model beats ViT baseline by 15.6% Dice","U-Net with Vision-LSTM and Chebyshev KAN nets 15.6% higher Dice","Vision-LSTM+KAN U-Net tops 3D organ Dice by 15.6%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the published UNETR and other comparison scores were obtained under the same preprocessing, training, and evaluation protocol as UNETVL, so the reported Dice gains reflect the architecture rather than the pipeline.","fun_headline_variants_meta":{"raw":{"variants":["UNetVL lifts 3D organ Dice up to 15.6% over UNETR","Chebyshev-KAN Vision-LSTM model beats ViT baseline by 15.6% Dice","U-Net with Vision-LSTM and Chebyshev KAN nets 15.6% higher Dice","Vision-LSTM+KAN U-Net tops 3D organ Dice by 15.6%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000708,"raw_usage":{"total_tokens":3191,"prompt_tokens":949,"completion_tokens":2242,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":2138}},"tokens_in":565,"tokens_out":2242,"duration_ms":15070,"temperature":1.0,"reasoning_tokens":2138,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:49:15.262548+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the UNETR baseline and UNETVL with identical data splits, preprocessing, augmentation, batch size, optimizer, and epoch counts, then compare per-fold mean Dice with standard deviations. If the reported 7.3% (ACDC) and 15.6% (AMOS) gaps shrink to roughly 2% or less, the claimed advantage is a protocol artifact rather than an architectural effect.","supporting_citations":[{"cited_title":"UNetVL: Enhancing 3D Medical Image Segmentation with Chebyshev KAN Powered Vision-LSTM","cited_arxiv_id":"2501.07017","evidence_quote":"It supplies the UNETR baseline Dice values and the encoder-decoder skeleton that UNETVL modifies."},{"cited_title":"As shown in Table 1, when comparing the performance of the ViL architecture with Chebyshev KAN to the UNETR baseline, a signifi- cant improvement is observed","cited_arxiv_id":null,"evidence_quote":"It supplies the Vision-LSTM encoder blocks, built from bidirectional mLSTM layers, that replace the ViT encoder."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It introduces the Kolmogorov-Arnold Network idea that Chebyshev KAN specializes."},{"cited_title":"[6] and Ji et al","cited_arxiv_id":null,"evidence_quote":"It provides the ACDC cardiac MRI benchmark and its manual anatomical labels used for evaluation."}],"review_version":1}