{"id":"7fb64182-5607-49b8-9e98-8ed358c81218","arxiv_id":"2507.16122","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"MLRU++ reports small Dice gains over UNETR++ on three public benchmarks using a lighter attention module and multiscale depthwise blocks.","lead":"MLRU++ is a lightweight 3D medical image segmentation network built on UNETR++, adding a channel-spatial attention module and a multiscale depthwise block. On Synapse, ACDC, and Lung benchmarks it reports average Dice scores of 87.57%, 93.00%, and 81.12%, while cutting parameter count.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported headline accuracy and efficiency are internally inconsistent: the same full model scores 87.57% in Table 1 and 87.85% in Table 5, and MLRU++ (46.09M params) is larger than UNETR++ (42.96M), contradicting the reduced-parameter claim.","rationale":"The reader's rejection is well supported, and my independent pass finds an even more direct problem than baseline comparability: the paper's own numbers for the proposed model disagree. Table 1 reports 87.57% for full MLRU++ on Synapse; Table 5 reports 87.85% for the same configuration; the abstract uses 87.57%. This is the central accuracy claim, and no training logs or code outputs are supplied to adjudicate. The efficiency claim is contradicted by the same tables: MLRU++ has 46.09M parameters versus UNETR++'s 42.96M, while the abstract and introduction claim significant parameter reduction. The ablation rows in Table 5 are internally mislabeled relative to Table 1, and the parameter jump from 44.92M to 46.09M when adding M2B is difficult to reconcile with Section 3.2's O(C) parameter claim at C=256. Appendix A.3 also contradicts the main text's 7x7 depthwise spatial attention. These are not external or consensus disagreements; they are internal inconsistencies in the reported evidence. Consequently, the central SOTA-and-lightweight assertion is not currently demonstrated. If the released code reproduces 87.57% and 46.09M, much of the accuracy/efficiency concern is resolved, but the manuscript as submitted does not provide that verification. I partially agree with the reader's weakest assumption: baseline comparability is also a real issue because baselines were not rerun under a shared protocol, but the internal inconsistency is more load-bearing because it does not depend on any external numbers. The architectural ideas (LCBAM, M2B) may be plausible, and the code availability is a positive signal, but the presented evidence does not support the paper's central claim as written.","tokens_in":13289,"tokens_out":3753,"duration_ms":36056,"concrete_test":"Run the released training code at https://github.com/1027865/MLRUPP on the Synapse dataset with the Section 4.2 settings (patch size 64x128x128, SGD lr=0.01, Nesterov momentum 0.99, deep-supervision weights [0.57, 0.29, 0.14]) and record the final average DSC and total parameter count for the full MLRU++ model. If the reproduced DSC is neither 87.57% nor 87.85%, or the parameter count deviates from 46.09M, the headline claim is unsupported. As a second check, instantiate the M2B block alone at C=256 and measure its parameter growth to test the O(C) assertion in Section 3.2.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that MLRU++ achieves state-of-the-art accuracy while reducing parameter count and computational cost. The paper's own tables undercut both parts of this claim. First, the headline accuracy is not pinned down: Table 1 reports 87.57% average DSC for full MLRU++ on Synapse, while Table 5 reports 87.85% for the same configuration (\"MLRU++ (LCBAM + M2B)\"), and the abstract uses 87.57%. No explanation is given for the discrepancy, and it is exactly the number on which the SOTA claim rests. Second, the efficiency claim is contradicted by the parameter tables: Table 1 gives MLRU++ 46.09M parameters versus UNETR++'s 42.96M, and Table 5 confirms 46.09M, yet the abstract and introduction say MLRU++ \"significantly reduc[es] parameter count.\" The ablation rows in Table 5 are also mismatched with Table 1: the row labeled \"LCBAM Only Encoder + Decoder\" has 86.36% DSC and 44.60M params, while Table 1's \"Encoder LCBAM only\" row has 86.36% but 44.92M params. Section 3.2 claims M2B adds O(C) parameters, but the jump from 44.92M to 46.09M is roughly 1.17M, which at C=256 is more consistent with a pointwise C-to-C projection than with an O(C) block. Appendix A.3 says LCBAM spatial attention uses depthwise 3x3 or 1x1 convolutions, whereas the main text Eq. (2) specifies a depthwise 7x7. These inconsistencies mean the evidence presented cannot establish either the accuracy or the efficiency half of the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MLRU++, a 3D medical image segmentation architecture built on a UNETR++-style encoder-decoder with two new components: LCBAM, a lightweight channel-and-spatial attention module, and M2B, a multiscale decoder block. Experiments are reported on Synapse, BTCV, ACDC, and Decathlon Lung, with claims of state-of-the-art Dice scores and reduced parameter count and computational cost. The authors also provide an ablation study and release source code.","tokens_in":13727,"tokens_out":6373,"duration_ms":57247,"significance":"If substantiated, a lightweight high-accuracy 3D segmentation architecture would be practically valuable, and the paper's attention to computational cost is commendable. The availability of source code and the inclusion of multi-dataset ablations are also positive. However, the manuscript currently contains multiple internal inconsistencies in the reported accuracy, parameter counts, and FLOPs, and the efficiency claim is contradicted by the paper's own tables. These issues prevent the reader from verifying either half of the central claim.","major_comments":[{"comment":"The headline accuracy for the full MLRU++ model is not reproducible from the paper's own tables: Table 1 reports an average DSC of 87.57% for MLRU++, Table 5 reports 87.85% for the same configuration 'MLRU++ (LCBAM + M2B)', and the abstract uses 87.57%. Since the state-of-the-art claim rests on this number, the discrepancy must be resolved with a single consistent experimental result.","section":"Abstract; Tables 1 and 5"},{"comment":"The efficiency claim is contradicted by the reported complexity numbers. Table 1 lists MLRU++ at 46.09M parameters versus UNETR++ at 42.96M, and Table 5 reports 66.06G FLOPs for MLRU++ versus 47.98G for UNETR++. The abstract's statement that MLRU++ 'significantly reduc[es] parameter count and computational cost' is therefore not supported by the evidence presented. The claim of the 'lowest HD95 of 7.53' in §5.1 is also inaccurate because Table 1 shows UNETR++ with the same value.","section":"Abstract; Table 1; Table 5"},{"comment":"The ablation results are internally inconsistent. Table 5 lists 'MLRU++ (LCBAM Only Encoder + Decoder)' with 86.36% DSC and 44.60M parameters, while 'MLRU++ (LCBAM Only in Encoder)' has 86.66% and 44.92M. Table 1's 'MLRU++ (Encoder LCBAM only)' row reports 86.36% and 44.92M, matching neither Table 5 row in both values. The text in §6.1 states that the encoder-plus-decoder variant (without M2B) 'performs better (86.66%)', which corresponds to the encoder-only row in Table 5. These mismatches make it impossible to determine the contribution of each component.","section":"§6.1; Table 5; Table 1"},{"comment":"The comparison protocol is not established. The paper states that 'all training configurations follow the setup in [31]' but lists patch sizes of 64×128×128 for Synapse, 32×192×192 for Lung, and 16×160×160 for ACDC, which differ from the configuration in [31]; moreover, none of the baseline models are re-trained under the authors' protocol. Since data splits, preprocessing, and training details are not controlled, the reported differences against cited baseline numbers (e.g., the 5.38% and 2.12% improvements claimed in the abstract) cannot be verified as genuine gains.","section":"§4.2; Tables 1-4"},{"comment":"The kernel-size specifications are inconsistent across the paper. Equation (2) defines spatial attention with a depthwise 7×7 convolution, Appendix A.3 states that LCBAM uses 'either depthwise 3×3 or 1×1 convolutions,' and §3.3 defines the MSDC kernel set as K={3,5,7} while Appendix B.1 uses K={1,3,5}. The parameter and FLOPs estimates in the appendix depend on these kernel sizes, so the efficiency analysis must be tied to the actual implementation.","section":"Eq. (2); §3.3; Appendix A.3; Appendix B.1"}],"minor_comments":[{"comment":"The column header 'Aotra' should be 'Aorta'.","section":"Table 1"},{"comment":"The text contains the typo 'visualizations0' instead of 'visualizations'.","section":"§5.4"},{"comment":"The name M2B is inconsistent: the abstract calls it 'Multiscale Bottleneck Block', Figure 1 calls it 'Multiscale Modulation Block', and §3.2 calls it 'MobileNet-Based Multiscale Block'. Please unify the terminology.","section":"Abstract; Figure 1; §3.2"},{"comment":"The claim that 'MLRU++ not only improves accuracy but also maintains competitive model complexity' is hard to evaluate because FLOPs are only reported in Table 5, not in the main comparison tables.","section":"§6.1"},{"comment":"The statement that MLRU++ achieves the 'lowest HD95 of 7.53' is not accurate because UNETR++ also achieves 7.53 in Table 1; the text should describe this as a tie or provide a different boundary metric.","section":"§5.1"}],"recommendation":"reject","confidential_remarks":"The paper's central claims are undermined by internal inconsistencies in the reported accuracy and efficiency numbers. Most critically, the parameter counts in Tables 1 and 5 show MLRU++ is larger than UNETR++, which directly contradicts the abstract's reduced-parameter claim. The ablation table and its accompanying text also disagree in several places. These are not superficial typos; they affect the validity of the main contribution. A revision would require re-running the experiments under a consistent protocol, re-training baselines, and substantially rewriting the claims. Given the scope of the required changes, I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"MLRU++ is a UNETR++ variant with a lightweight channel–spatial attention module (LCBAM) and a MobileNet-style multiscale depthwise block (M2B) in the decoder. The components are known techniques—1×1/depthwise convolutions, channel shuffle, multiscale fusion—but the specific combination for 3D segmentation is new, and the authors release code and evaluate on four public benchmarks. That part is legitimate.\n\nThe problem is that the paper's own tables undercut both halves of the central claim (SOTA accuracy + fewer parameters). The same full model scores 87.57% in Table 1 and 87.85% in Table 5. The parameter count is 46.09M vs UNETR++'s 42.96M, yet the abstract claims 'significantly reducing parameter count.' The ablation text says the LCBAM-in-both-encoder-and-decoder variant (without M2B) hits 86.66%, but Table 5 lists 86.36% for that variant and 86.66% for encoder-only. The main text specifies a depthwise 7×7 in spatial attention; the appendix says 3×3 or 1×1. MSDC kernel sets differ between {3,5,7} and {1,3,5}. The abstract's 5.38% and 2.12% improvements over leading models don't appear in any table—the gaps vs UNETR++ on Synapse and ACDC are 0.35 and 0.17 points. There are also no error bars, and baselines are not re-run under a common protocol, despite the claim that training follows [31] with different patch sizes.\n\nThese aren't cosmetic typos; they hit the two sentences the paper is selling. As written, the evidence doesn't establish either state-of-the-art accuracy or reduced complexity. That said, the architecture isn't incoherent and the code is public, so a careful revision could salvage a modest incremental contribution. I'd send it to review—a serious referee should see it—but my recommendation would be reject/major revision in current form.","headline":"A sensible but unverified incremental architecture whose own tables contradict the SOTA-accuracy and reduced-parameter claims—needs major revision before the results can be trusted.","tokens_in":14316,"tokens_out":3278,"would_cite":false,"duration_ms":31553,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MLRU++ claims to beat current state-of-the-art 3D medical image segmentation models on four benchmarks while using fewer parameters.","keywords":["3D medical image segmentation","lightweight attention","depthwise separable convolution","multiscale feature aggregation","UNETR++","Dice score","channel-spatial attention","residual encoder-decoder"],"falsifier":"Re-run the comparison by training UNETR++, nnFormer, and nnUNet on the exact same data splits, patch sizes, augmentations, and loss used for MLRU++, then compare average Dice on Synapse, ACDC, and Lung; if the gaps drop below statistical significance, the state-of-the-art claim is not supported.","tokens_in":13076,"feed_emoji":"🩻","tokens_out":3860,"duration_ms":36065,"temperature":0.7,"pith_summary":"This paper proposes MLRU++, a 3D medical image segmentation network built on a UNETR++-style encoder-decoder but made lighter and more accurate. The two novel pieces are LCBAM, a channel-and-spatial attention module that replaces CBAM's MLP and full convolutions with 1x1 and depthwise convolutions, and M2B, a MobileNetV2-inspired decoder block with parallel depthwise convolutions at multiple kernel sizes plus a channel shuffle. On four public benchmarks (Synapse, BTCV, ACDC, Decathlon Lung), the authors report the highest average Dice among the compared models, with gains over UNETR++ of 5.38 points on Synapse and 2.12 points on ACDC, while keeping parameter count and FLOPs low. The claim is that this hybrid design gives both better accuracy and better efficiency than current state-of-the-art 3D segmentation models.","feed_headline":"Lighter 3D segmentation model beats UNETR++ on four benchmarks","feed_subtitle":"Two lightweight modules—channel-spatial attention and multiscale depthwise blocks—raise Dice while cutting parameters.","key_machinery":"The load-bearing components are LCBAM and M2B. LCBAM computes channel attention as two 1x1 convolutions after adaptive average pooling, and spatial attention as a 7x7 depthwise convolution on pooled features, giving roughly half the channel-attention parameters of CBAM and about 90% fewer spatial-attention FLOPs. M2B applies parallel depthwise convolutions with kernel sizes 3, 5, and 7, sums them with a residual connection, and shuffles channels before a pointwise projection, adding only O(C) parameters. These modules sit inside a residual UNETR++ encoder-decoder with deep supervision, and they carry the argument that multiscale context can be added without quadratic attention cost.","core_discovery":"On the paper's own terms, the central discovery is that a lightweight residual UNETR++ architecture can beat larger hybrid CNN-Transformer models on four volumetric benchmarks. The authors state that MLRU++ reaches average Dice scores of 87.57% on Synapse, 85.20% on BTCV, 93.00% on ACDC, and 81.12% on the Decathlon Lung dataset, exceeding UNETR++, nnFormer, and nnUNet on most organs while using a comparable or smaller parameter budget. Ablations attribute the gain to the combination of LCBAM and M2B: adding the multiscale block raises Synapse Dice from 86.66% to 87.85% in the ablation table, and the full model outperforms its LCBAM-only variant on ACDC as well.","pith_inferences":["Because the baselines were not re-trained under the authors' protocol, the exact size of the claimed accuracy advantage is an open question; a shared-protocol benchmark would settle it.","The M2B design with 3, 5, and 7 depthwise kernels suggests the module's benefit should scale with the range of anatomical structure sizes in the target organ, so one testable extension is applying it to datasets with very large and very small structures together.","The same lightweight-attention recipe could be applied to other dense prediction tasks such as 3D detection or image restoration, where the FLOPs budget is similarly constrained.","Nothing in the paper measures inference latency or peak GPU memory on clinical hardware, so reporting wall-clock time and memory usage would make the efficiency claim more concrete."],"forward_implications":["MLRU++ can be trained on a single 32 GB V100 GPU for 3D segmentation, pointing to feasible use in clinical settings with limited compute.","The per-organ gains on Synapse are concentrated in small structures such as the gallbladder and pancreas, suggesting the multiscale block helps where boundary detail matters.","The parameter and FLOPs reductions of LCBAM relative to CBAM imply that attention can be added to volumetric models without quadratic cost, which could transfer to other 3D vision tasks.","Deep supervision with compound loss weights is part of the design, and the reported Dice improvements on two independent datasets indicate the supervision strategy generalizes across modalities."],"supporting_citations":[{"why":"Defines the UNETR++ architecture and baseline scores that MLRU++ extends and compares against.","marker":"[25]"},{"why":"Supplies the training configuration that the paper claims to follow for fair comparison.","marker":"[31]"},{"why":"Defines CBAM, the attention module that LCBAM replaces and whose efficiency it is measured against.","marker":"[28]"},{"why":"Introduces the MobileNetV2 inverted-residual design that the M2B block is based on.","marker":"[24]"},{"why":"Provides nnUNet, a strong baseline whose Dice scores MLRU++ claims to surpass on all four datasets.","marker":"[14]"},{"why":"Provides the UNETR baseline used in ablations and as a key comparison point.","marker":"[9]"}],"fun_headline_variants":["Lightweight 3D segmentation beats heavier models on four benchmarks","Two lightweight modules lift Dice and cut parameters in 3D segmentation","Slim UNETR++ variant outperforms bigger models on four datasets","Efficient attention and multiscale blocks give lighter 3D segmentation","MLRU++: lightweight residual UNETR++ with attention beats baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains assume that the baseline scores taken from earlier papers were measured under conditions equivalent to MLRU++'s training setup, since the baselines were not re-run in this study.","fun_headline_variants_meta":{"raw":{"variants":["Lightweight 3D segmentation beats heavier models on four benchmarks","Two lightweight modules lift Dice and cut parameters in 3D segmentation","Slim UNETR++ variant outperforms bigger models on four datasets","Efficient attention and multiscale blocks give lighter 3D segmentation","MLRU++: lightweight residual UNETR++ with attention beats baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000815,"raw_usage":{"total_tokens":3595,"prompt_tokens":994,"completion_tokens":2601,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":2507}},"tokens_in":610,"tokens_out":2601,"duration_ms":18287,"temperature":1.0,"reasoning_tokens":2507,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:17:23.452026+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the comparison by training UNETR++, nnFormer, and nnUNet on the exact same data splits, patch sizes, augmentations, and loss used for MLRU++, then compare average Dice on Synapse, ACDC, and Lung; if the gaps drop below statistical significance, the state-of-the-art claim is not supported.","supporting_citations":[{"cited_title":"In: International MICCAI brainlesion workshop","cited_arxiv_id":null,"evidence_quote":"Provides the UNETR baseline used in ablations and as a key comparison point."},{"cited_title":"In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018)","cited_arxiv_id":null,"evidence_quote":"Defines the UNETR++ architecture and baseline scores that MLRU++ extends and compares against."},{"cited_title":"In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)","cited_arxiv_id":null,"evidence_quote":"Supplies the training configuration that the paper claims to follow for fair comparison."},{"cited_title":"Medical Image Analysis 73, 102222 (2022)","cited_arxiv_id":null,"evidence_quote":"Defines CBAM, the attention module that LCBAM replaces and whose efficiency it is measured against."},{"cited_title":"Bounds on the number of maximal subgroups of finite groups","cited_arxiv_id":"2203.05597","evidence_quote":"Introduces the MobileNetV2 inverted-residual design that the M2B block is based on."},{"cited_title":"In: Proc","cited_arxiv_id":null,"evidence_quote":"Provides nnUNet, a strong baseline whose Dice scores MLRU++ claims to surpass on all four datasets."}],"review_version":1}