{"id":"3b4606c5-2f86-4797-bc0e-4dc7bebec480","arxiv_id":"2412.06262","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Three plug-and-play decoders built from discretized neural memory ODEs shrink U-like segmentation networks' parameters and FLOPs while roughly preserving accuracy on three skin lesion datasets.","lead":"The authors replace the decoder half of U-shaped medical image segmentation networks with a small module built from discretized ordinary differential equations, cutting parameters by roughly a fifth to a half and FLOPs by up to 74% on skin lesion datasets. The module is designed to plug into any U-like network without redesigning the rest of the architecture.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Performance-parity claim rests on comparing modified networks to external baselines from prior papers with different data splits and training recipes; matched re-training is required to confirm that accuracy is maintained.","rationale":"The parameter and FLOP reductions are arithmetic and backed by public code; they are not in serious doubt. The load-bearing part of the central claim is that these reductions come without sacrificing accuracy ('while maintaining performance'). That claim rests entirely on comparisons in Table 1 against numbers imported from earlier papers with heterogeneous experimental setups. Because the modified networks were trained under the authors' own pipeline with a random 7:3 split, any observed improvement or parity could be an artifact of training conditions rather than the decoder design. This is the classic weak-comparison problem: without matched baselines, the central claim is unverified. The ablation in Table 2 provides only a partial hint at a matched comparison for UNet on PH2, but its consistency with Table 1 is unclear (mIoU matches the cited value while DSC differs), so it cannot rescue the main table. A single controlled re-training experiment, e.g., Att-UNet versus Att-UNet(EED) on PH2 under identical conditions, would settle the matter. The reader's CONDITIONAL verdict is appropriate; the efficiency result can stand, but the accuracy-parity claim should be conditioned on matched baselines.","tokens_in":12256,"tokens_out":9164,"duration_ms":87412,"concrete_test":"Retrain all five original baseline networks (UNet, Att-UNet, MALUNet, EGE-UNet, UNeXt) using the authors' exact pipeline of Section 4.2: same random 7:3 split seed, resizing to 256x256, augmentation, AdamW, CosineAnnealingLR, 300 epochs, batch size 8, and evaluate with the same mIoU/DSC code. The decisive minimal check is Att-UNet on PH2: if the retrained original Att-UNet reaches or exceeds the 0.905 mIoU of Att-UNet(EED), the claim that the decoder preserves accuracy is falsified. This test removes the confound of split/training recipe and directly settles whether the reported performance parity is real.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the nmODE decoders 'maintain performance' while reducing parameters/FLOPs is supported in Table 1 by comparing modified networks against baseline metrics copied from Cheng et al. (2023), Wu et al. (2022), and Ruan et al. (2023), as stated in the Table 1 footnote. Section 4.2 specifies that all modified networks were trained on a random 7:3 split of each dataset with AdamW, CosineAnnealingLR, 300 epochs, batch size 8, and augmentation. There is no evidence the cited baselines used the same split, resolution, training schedule, or metric protocol. For example, Att-UNet(EED) reports mIoU 0.905 on PH2 versus the cited Att-UNet 0.858, a 4.7-point gain that could plausibly stem from a more favorable split or longer training rather than from the decoder. The ambiguity is heightened by Table 2, where the 'origin' UNet row has mIoU 0.8780 but DSC 0.9342, while the cited UNet in Table 1 has DSC 0.919; if the authors actually trained the original UNet in the ablation, the same matched baseline should have been used in Table 1. Until the original networks are retrained under the identical pipeline, the 'maintaining performance' assertion is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes replacing the decoder of U-shaped segmentation networks with lightweight blocks obtained by discretizing neural memory ordinary differential equations (nmODEs). Three discretizations are used: explicit Euler (EED), Heun (HD), and linear multistep (LMD). The upward path of the decoder is intended to be parameter-free, with trainable g-functions on the skip connections performing channel and spatial alignment. Experiments on PH2, ISIC2017, and ISIC2018 embed these decoders into UNet, Att-UNet, MALUNet, EGE-UNet, and UNeXt, reporting roughly 20–50% parameter reductions and up to 74% FLOP reductions while claiming that segmentation accuracy is maintained or slightly improved. Additional ablations on UNet/PH2 study the effect of the number of channels in the initial state y(0).","tokens_in":12513,"tokens_out":6345,"duration_ms":61673,"significance":"If the performance-parity and efficiency claims hold, the proposed decoders would be a practically useful plug-and-play way to compress existing U-like networks while preserving accuracy. The derivations of the three discretized blocks are explicit, the parameter and FLOP arithmetic in Tables 1–3 is transparent, and the ablation study in Table 2 shows internal consistency in isolating the decoder's contribution. The work is, however, largely an application of the previously published nmODE framework (Yi 2023; He et al. 2023; Wang et al. 2024) rather than a new theoretical construction. The main value rests on the empirical efficiency claim, which is currently weakened by the inconsistent baselines described in the major comments.","major_comments":[{"comment":"The central claim that segmentation performance is maintained is not supported by the comparisons as reported. Modified networks were trained by the authors on a random 7:3 split with AdamW, CosineAnnealingLR, 300 epochs, batch size 8, and augmentation (§4.2), while most baseline rows are copied from Cheng et al. (2023), Wu et al. (2022), and Ruan et al. (2023), as the Table 1 footnote states. Those baselines need not use the same split, resolution, training schedule, or metric protocol. The discrepancy is visible within the paper: Table 1 lists UNet with DSC 0.919, whereas the authors' own “origin” UNet in Table 2 has DSC 0.9342 under the same input resolution. Likewise, Att-UNet(EED) exceeds the cited Att-UNet by 4.7 mIoU points on PH2, a margin large enough to be attributable to training setup rather than the decoder. The authors should retrain each original network under their own pipeline and compare matched pairs, or explicitly report the original networks' results under identical splits and settings.","section":"§4.3, Table 1"},{"comment":"The ablation table reports “origin” UNet with mIoU 0.8780 and DSC 0.9342, while Table 1 cites UNet with mIoU 0.878 and DSC 0.919. The near match in mIoU but clear mismatch in DSC indicates that the external baseline and the ablation baseline are not the same model or run. Table 2 therefore cannot serve as a matched control for the efficiency comparisons in Table 1. This internal inconsistency should be resolved by using one consistent UNet baseline across all tables.","section":"§4.5, Table 2"},{"comment":"The manuscript states that the upward path is parameter-free and that the f-function contains only batch normalization, which is described as a parameter-free operation. In standard PyTorch, BatchNorm2d has learnable weight and bias by default (affine=True). If the implementation uses affine=False, this should be stated explicitly and reflected in the code and parameter counts; if not, the statement that the decoder has no parameters on the upward path is inaccurate and the parameter savings in Tables 1–3 need to be recomputed or clarified to show they are unaffected.","section":"§3.2, Function Selection & Initial Value Problem"}],"minor_comments":[{"comment":"Equation (7) refers to ẏl without defining how this discrete derivative is computed inside the Heun block. From the derivation it appears to be the explicit-Euler intermediate value, but this should be stated in the text and labeled in Fig. 4(b).","section":"§3.2, Eq. (7)"},{"comment":"The notation “xn+1.θn+1” in Eq. (6) uses a period instead of a comma, and the index convention shifts to xl−1 and θl−1 in Eq. (7). Please align the indices so the reader can map the derivation onto the network layers.","section":"§3.2, Eq. (6) and Eq. (7)"},{"comment":"The text says CosineAnnealingLR is used “setting the maximum iteration count to 50” while training spans 300 epochs; these numbers are inconsistent. Please specify whether T_max is 50 epochs, 50 iterations, or something else.","section":"§4.2"},{"comment":"The promised 20–50% parameter reduction and up to 74% FLOP reduction are presented as general statements, but they depend on comparisons with external baselines. A sentence noting that these figures use the reported baseline sources, and that matched retraining is needed for strict comparison, would make the claim more precise.","section":"Abstract and §4.3"},{"comment":"Standard deviations are reported for the authors' models only. For visual fairness, either report mean±std for all baselines or state clearly that baseline values are single numbers taken from the original publications.","section":"Table 1"},{"comment":"The observation that HD underperforms EED and LMD despite having more parameters is mentioned but not discussed. Since this is surprising for a higher-order method, a brief explanation would help practitioners choose among the three decoders.","section":"§4.3"},{"comment":"The caption says “the orange yl and red yl−1 are used as the upward path inputs and outputs,” but the diagram does not clearly identify which arrows are orange or red. Please add explicit labels in the figure.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The paper's novelty is mainly an application of the authors' previously published nmODE framework, with a substantial amount of self-citation. That is not by itself disqualifying, but the editor may want to consider whether the contribution is sufficiently incremental for the target venue. The Table 1 versus Table 2 baseline discrepancy should be fixed before any acceptance decision, and the authors should either share the retrained baseline checkpoints or provide enough detail for exact reproduction of the matched comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a reasonable architecture-engineering paper with a plausible efficiency result, but the accuracy-parity claim is not yet backed by matched comparisons. The parameter and FLOP reductions are easy to verify and the ablation on PH2 is a good step. However, the headline comparisons in Table 1 compare modified networks to numbers copied from three other papers, with different splits and training schedules. The stress-test note is on target. The UNet row alone shows the problem: DSC in Table 1 is 0.919, while the 'origin' UNet in Table 2 (same paper, same dataset, presumably their own run) is 0.9342. That inconsistency needs explanation; if the authors trained the original UNet, they should have used it as the baseline in Table 1 for all networks they actually ran.\n\nWhat is genuinely new: the idea of using Euler, Heun, and linear-multistep discretizations as parameter-free decoder replacements across five U-like networks, with shared weights on the upward path. The numerical methods are textbook, but the architectural packaging is a legitimate engineering contribution, and the code is public, which is a real plus. The ablations on y(0) channel count and the simplified decoder are informative.\n\nSoft spots, in order: (1) matched baselines are missing for most comparisons; only UNet on PH2 has anything close, and that has the DSC discrepancy. (2) The paper reports each network with one decoder variant, apparently selected post hoc; no systematic comparison of all three decoders per network. (3) 'All U-like networks' is a stretch from five architectures, three of which are from the same group. (4) The state-of-the-art claims on PH2 are inflated because the comparison set is small.\n\nOn balance, the efficiency claim (20-50% parameters, up to 74% FLOPs) is credible and reproducible. The accuracy-parity claim needs matched re-training before it can be accepted. This deserves a serious referee, but the revision requirements are substantial: retrain or cite matching baselines, report the DSC inconsistency, and tone down the generality claim. I would send it to review, but I would not cite it in my own work until the baseline issue is resolved.","headline":"Efficiency numbers are credible, but the performance-parity claim rests on uncontrolled external baselines until matched re-training confirms it.","tokens_in":13059,"tokens_out":2568,"would_cite":false,"duration_ms":24170,"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 proposes three plug-and-play decoders obtained by discretizing neural memory ordinary differential equations with explicit Euler, Heun, and linear multistep methods, and claims they replace the decoders of U-like networks to cut…","keywords":["neural memory ordinary differential equations","medical image segmentation","lightweight decoder","U-like networks","discretized ODE solvers","skin lesion segmentation","parameter reduction","plug-and-play decoder"],"falsifier":"Retrain one of the original networks (for example EGE-UNet on ISIC2018) under the exact same 7:3 random split, augmentation, optimizer, and 300-epoch schedule used for the modified decoders, and compare mIoU and Dice; if the retrained baseline matches or beats the nmODE-decoder version, the claim that accuracy is maintained loses its support.","tokens_in":12056,"feed_emoji":"🩺","tokens_out":8591,"duration_ms":71636,"temperature":0.7,"pith_summary":"This paper tries to establish that the decoder half of a U-like segmentation network can be replaced by a small, plug-in module derived from a neural memory ordinary differential equation (nmODE) without hurting accuracy and with large efficiency gains. Three discretizations of the nmODE—explicit Euler, Heun, and a linear multistep method—give three decoder variants (EED, HD, LMD). Across the PH2, ISIC2017, and ISIC2018 skin-lesion datasets, embedding these decoders in UNet, Att-UNet, UNeXt, MALUNet, and EGE-UNet reduces parameters by roughly 20–50% and FLOPs by up to 74%, while reported mIoU and Dice scores stay level or improve. If the pattern holds, the result matters because it offers a general way to slim existing U-like networks without redesigning their encoders or training from scratch.","feed_headline":"Swap-in ODE decoders cut U-net parameters by 20–50%","feed_subtitle":"Three plug-and-play decoder modules keep segmentation accuracy while cutting FLOPs by up to 74% on five U-like networks.","key_machinery":"The central object is the discretized nmODE decoder: a neural memory ordinary differential equation (an ODE whose state $y(t)$ has a zero initial value and receives features from skip connections as external input $x(t)$) used to define the upward path of a U-like network. The work it does is to replace the decoder's many learned convolutional layers with a small number of parameterized skip-connection transforms $g(x_l,\\theta_l)$ and a parameter-free state update $f(\\cdot)$; the three numerical solvers—explicit Euler, Heun, and linear multistep—turn the continuous equation into the concrete layer recurrences that make the module plug-and-play.","core_discovery":"On the paper's own terms, the discovery is that the decoder of a U-like network can be reformulated as an nmODE initial-value problem and solved by standard numerical integrators with a parameterless upward path. The nmODE state equation is $\\dot{y}(t) = -y(t) + f\\bigl(y(t) + g(x(t), \\theta_t)\\bigr)$, where skip-connection features $x(t)$ enter as sequential external inputs, the upward-path activations are the state $y(t)$, and $y(0)$ is initialized at zero instead of at the data. Discretizing this equation by explicit Euler, Heun, or linear multistep methods yields the layer recurrences $y_{l-1} = (1-\\delta)y_l + \\delta f(y_l + g(x_l,\\theta_l))$ (EED), a predictor-corrector form (HD), and $y_{l-1} = y_{l+1} - 2\\delta y_l + 2\\delta f(y_l + g(x_l,\\theta_l))$ (LMD). Because the $f$ function is parameter-free and the upward path shares weights, only the skip-connection $g$ functions carry parameters; this is what produces the 20–50% parameter reduction and up to 74% FLOP reduction reported on five U-like networks over three datasets.","pith_inferences":["If the parameterless upward path really transfers across architectures, the same decoder modules should also slim transformer-based U-shaped segmentation networks (for example Swin-UNet) without retraining their encoders; the paper only tests CNN- and MLP-based U-nets, so this is an extrapolation.","The ablation shows that the initial state's channel count is the main cost lever, so an obvious extension is an adaptive or grouped channel schedule that grows capacity only where the segmentation task needs it.","Heun's method, despite the extra g-function, was not consistently the best performer; this suggests the discretization's formal accuracy order may matter less than how skip-connection features are gated into the state, a hypothesis one could test by swapping the g-function design while holding the solver fixed.","Because the baselines come from other papers, the efficiency numbers are the solid part of the evidence while the accuracy-parity claim awaits a same-pipeline head-to-head; the authors' own ablations on UNet already show that the decoder algorithm, not just channel reduction, drives the gains."],"forward_implications":["Att-UNet with the Euler decoder (EED) on PH2 drops from 34.88M parameters and 66.63 GFLOPs to 18.91M and 17.34 GFLOPs while mIoU rises from 0.858 to 0.905.","UNet with the linear-multistep decoder (LMD) on PH2 cuts parameters from 13.40M to 9.45M and raises mIoU from 0.878 to 0.905.","EGE-UNet with LMD on ISIC2018 cuts parameters from 0.053M to 0.041M and raises mIoU from 0.809 to 0.813, reaching reported state-of-the-art.","MALUNet with EED on ISIC2018 cuts parameters from 0.177M to 0.098M with mIoU rising from 0.803 to 0.806.","The three decoder variants are drop-in replacements: they require no changes to the encoder and no re-devised lightweight backbone, supporting the paper's claim of adaptability across U-like architectures."],"supporting_citations":[{"why":"Introduces nmODEs and supplies the differential equation and memory mechanism that the three decoders discretize.","marker":"[Yi, 2023]"},{"why":"Establishes neural ODEs as parameter-efficient network layers, the design rationale for the lightweight upward path.","marker":"[Chen et al., 2018]"},{"why":"Defines the UNet encoder-decoder-skip-connection pattern and provides the original UNet baseline used in experiments.","marker":"[Ronneberger et al., 2015]"},{"why":"Provides the explicit Euler discretization underlying the EED decoder and its layer recurrence.","marker":"[Euler, 1845]"},{"why":"Provides Heun's predictor-corrector method underlying the HD decoder.","marker":"[Heun, 1900]"},{"why":"Provides the linear multistep method underlying the LMD decoder's two-state recurrence.","marker":"[Bashforth and Adams, 1883]"},{"why":"Source of published PH2 baseline metric values that the modified networks are compared against.","marker":"[Cheng et al., 2023]"},{"why":"Source of published ISIC2017 and ISIC2018 baseline values and the EGE-UNet architecture modified in experiments.","marker":"[Ruan et al., 2023]"},{"why":"Defines Att-UNet, the architecture whose decoder is replaced and whose baseline improves on PH2.","marker":"[Oktay et al., 2018]"},{"why":"Defines UNeXt, the lightweight MLP-based architecture modified to show adaptability beyond convolutional U-nets.","marker":"[Valanarasu and Patel, 2022]"}],"fun_headline_variants":["Plug-and-play ODE decoders slim any U-net by 20–50%","Neural memory ODE decoders cut FLOPs up to 74%","Swap in nmODE decoders to shrink U-nets without accuracy loss","Lightweight U-net decoders: fewer parameters, up to 74% less compute","ODE-based decoders make U-nets leaner with 20–50% fewer parameters"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The performance claim assumes the baseline scores quoted from other papers were produced under training conditions equivalent enough to the authors' 7:3 split, 300-epoch schedule, and data augmentation that the comparison is fair.","fun_headline_variants_meta":{"raw":{"variants":["Plug-and-play ODE decoders slim any U-net by 20–50%","Neural memory ODE decoders cut FLOPs up to 74%","Swap in nmODE decoders to shrink U-nets without accuracy loss","Lightweight U-net decoders: fewer parameters, up to 74% less compute","ODE-based decoders make U-nets leaner with 20–50% fewer parameters"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000283,"raw_usage":{"total_tokens":1735,"prompt_tokens":1071,"completion_tokens":664,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":687,"completion_tokens_details":{"reasoning_tokens":556}},"tokens_in":687,"tokens_out":664,"duration_ms":5691,"temperature":1.0,"reasoning_tokens":556,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:51:07.424474+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain one of the original networks (for example EGE-UNet on ISIC2018) under the exact same 7:3 random split, augmentation, optimizer, and 300-epoch schedule used for the modified decoders, and compare mIoU and Dice; if the retrained baseline matches or beats the nmODE-decoder version, the claim that accuracy is maintained loses its support.","supporting_citations":[{"cited_title":"Neural ordinary dif- ferential equations","cited_arxiv_id":null,"evidence_quote":"Establishes neural ODEs as parameter-efficient network layers, the design rationale for the lightweight upward path."},{"cited_title":"Neue methoden zur approxi- mativen integration der differentialgleichungen einer un- abh¨angigen ver ¨anderlichen","cited_arxiv_id":null,"evidence_quote":"Provides Heun's predictor-corrector method underlying the HD decoder."},{"cited_title":"Segnetr: Rethinking the local-global interactions and skip connections in u-shaped networks,","cited_arxiv_id":null,"evidence_quote":"Source of published PH2 baseline metric values that the modified networks are compared against."},{"cited_title":"EGE-UNet: an Efficient Group Enhanced UNet for skin lesion segmentation","cited_arxiv_id":"2307.08473","evidence_quote":"Source of published ISIC2017 and ISIC2018 baseline values and the EGE-UNet architecture modified in experiments."},{"cited_title":"At- tention u-net: Learning where to look for the pancreas,","cited_arxiv_id":null,"evidence_quote":"Defines Att-UNet, the architecture whose decoder is replaced and whose baseline improves on PH2."},{"cited_title":"Unext: Mlp-based rapid medical im- age segmentation network","cited_arxiv_id":null,"evidence_quote":"Defines UNeXt, the lightweight MLP-based architecture modified to show adaptability beyond convolutional U-nets."}],"review_version":1}