{"id":"082421f7-dbca-4428-9067-47e4ed4455f1","arxiv_id":"1908.08466","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A sigmoid-blended instance-layer normalization with an added group normalization step reports modest Dice gains for U-Net cardiac MRI segmentation in five of six cross-validation folds.","lead":"This paper proposes a new normalization layer for U-Net-style image segmentation that mixes instance and layer statistics with a learned, smooth weight and then re-normalizes the result. On two cardiac MRI datasets it reports slightly higher average Dice scores than standard normalizations in most test folds, but the gains are small and not statistically tested.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed ILN accuracy gains rest on single two-epoch runs per method, with per-subject standard deviations several times larger than the mean differences; the 1–2% DSC advantage may be training noise rather than a real method effect.","rationale":"The reader's weakest_assumption correctly identifies the core problem: a single two-epoch run per method, with per-subject standard deviations much larger than the mean differences, cannot support a reliable ranking of normalization layers. My stress-test confirms this and adds the specific numbers from Table 3 and the learning-rate-selection protocol from Section 2.2. No additional load-bearing flaw emerged: the ILN construction is mathematically coherent, the comparison to Clip/Softmax in Section 3.1 is internally consistent, and the authors are appropriately cautious in noting the LV-3 exception and the lack of ground truth for ρ. The most important unresolved issue remains the statistical reliability of the headline improvement, which the paper does not address. Therefore the conditional verdict is appropriate; the paper should not be rejected outright because the method is simple, plausible, and the experimental design is transparently described, but it should not be accepted as establishing 'noticeable accuracy improvements' without repeated runs and significance testing.","tokens_in":6182,"tokens_out":1769,"duration_ms":19756,"concrete_test":"Re-run the Table 3 comparison with at least 10 random seeds per normalization method while holding the learning-rate selection protocol fixed, and report the mean and standard deviation of per-subject DSC across seeds for each of the six cross-validation folds. Then apply a paired test (e.g., Wilcoxon signed-rank on per-subject DSC, or a paired t-test on fold-level means) comparing ILN to IN. If ILN does not beat IN in at least five of the six folds with p<0.05, the claimed consistent accuracy improvement should be withdrawn.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that ILN 'outperforms previous traditional and popular normalization methods with noticeable accuracy improvements' and gives 'almost 2% DSC'—is supported only by Table 3. The mean ILN-minus-IN DSC differences are 0.005, 0.022, 0.025, 0.014, 0.017, and -0.009 across RV-1/2/3 and LV-1/2/3, while the reported per-subject standard deviations range from 0.098 to 0.323. Each tabulated entry comes from a single run: Section 2.2 states that 'two epochs were trained for each experiment' and that '[f]ive initial learning rates ... were tested for each experiment and the best value was selected to be shown.' This protocol has two problems. First, no repeated-run variance is reported, so the table's standard deviations reflect inter-subject variability, not the uncertainty in the training procedure; a different random initialization or data split could plausibly reverse the ranking given the size of the mean differences. Second, selecting the learning rate after seeing results, without a held-out validation set or multiple seeds, introduces selection bias and makes the 'best' result an upper bound rather than an unbiased estimate. Because the claimed improvement is only about one percentage point on average and the noise floor is an order of magnitude larger, the evidence does not establish that ILN is consistently better than IN, LN, or GN4. The method itself is plausible and the architecture is simple, but the evaluation is underpowered enough that the main quantitative conclusion is not supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Instance-Layer Normalization (ILN), which combines Instance Normalization and Layer Normalization feature maps using a Sigmoid-weighted trainable parameter, and then applies Group Normalization with 16 groups (GN16) to the combined feature map. The method is motivated by two perceived weaknesses of Batch-Instance Normalization: the non-differentiability of the Clip function at 0 and 1, and the loss of a normalized distribution after combining feature maps. The authors validate ILN on right-ventricle and left-ventricle cardiac MRI segmentation with a U-Net, comparing against no normalization, Instance Normalization, Layer Normalization, and Group Normalization over three-fold cross-validation. They report that ILN achieves the highest Dice Similarity Coefficient in five of six folds and claim an improvement of almost 2% DSC, concluding that ILN is an effective normalization strategy for medical image segmentation.","tokens_in":6492,"tokens_out":5883,"duration_ms":55731,"significance":"If the claimed accuracy gains were robust, ILN would be a useful and inexpensive drop-in normalization layer: it adds only one trainable parameter per layer plus a group-normalization step, and the manuscript tests it on public cardiac MRI benchmarks. A clear strength is that the authors re-implement all baselines in a common low-level TensorFlow framework, which is fairer than comparing against high-level API implementations. The paper is clearly written and the architectural idea is plausible. However, the current evidence is not strong enough to establish the central claim: the evaluation uses single two-epoch runs, post-hoc learning-rate selection, and mean differences between methods that are an order of magnitude smaller than the reported per-subject standard deviations. The authors honestly acknowledge one failure (LV-3) and that generalization to other tasks needs further validation, but the quantitative conclusion requires substantially stronger experimental support before the claimed improvement can be considered established.","major_comments":[{"comment":"The central claim that ILN outperforms IN, LN, and GN4 rests on a single two-epoch run per method with the best learning rate selected post hoc. The mean ILN-minus-IN differences in Table 3 are 0.005, 0.022, 0.025, 0.014, 0.017, and -0.009 across RV-1/2/3 and LV-1/2/3, while the reported per-subject standard deviations range from 0.098 to 0.323. Because no repeated-seed variance is reported and no significance test is applied, the observed differences are within the noise floor and do not establish a consistent improvement.","section":"Section 2.2 and Table 3"},{"comment":"Selecting the best of five initial learning rates from the test-fold results is a form of selection bias, making the reported values upper bounds rather than unbiased estimates of expected performance. The authors should either fix the learning rate a priori, tune it on a held-out validation split, or report the performance for all five learning rates and multiple random seeds.","section":"Section 2.2"},{"comment":"Training for only two epochs, with the learning rate divided by five at the second epoch, is an arbitrary early stopping point; rankings at epoch two may not reflect converged or practically relevant performance. The 'almost 2% DSC' claim in Section 5 is also not tied to a specific baseline: the mean ILN advantage over IN is about 1.2%, whereas over LN and GN4 it is about 2.3-2.8%, so the claimed improvement depends on the chosen comparison.","section":"Sections 2.2 and 5"},{"comment":"The experiment intended to validate Sigmoid over Clip does not include the original BIN method: the Clip row combines IN and LN with a single clipped weight, whereas BIN combines BN and IN. Consequently, the paper does not directly demonstrate an improvement over the actual BIN method it criticizes in the introduction.","section":"Section 3.1 and Table 1"}],"minor_comments":[{"comment":"The text states that '// is exact division', but in common programming languages '//' denotes floor division; if the channel count C is not a multiple of 16, the grouping formula with M = C//16 leaves some channels unnormalized or misassigned. Please clarify the intended grouping for arbitrary C.","section":"Equations (5) and (6)"},{"comment":"Since ILN contains GN16 as a component, the main comparison table should include a GN16-alone row rather than relying on the statement that GN16 'performed similarly to IN [16]'; without this row, the reader cannot separate the contribution of the Sigmoid combination from the contribution of the added GN16 step.","section":"Table 3"},{"comment":"The training curves of ρ in Figure 3 are presented without quantitative analysis, and the text acknowledges that the correctness of the curves cannot be judged; consider reporting the distribution or final values of ρ across layers, or relating them to layer depth, to make the figure informative.","section":"Section 3.5"},{"comment":"The claim that the Clip function 'is not differentiable at input values of 0 and 1' is true only at two isolated points, and subgradients are standard in deep learning; the paper should state why this non-differentiability is practically harmful rather than merely theoretically present.","section":"Section 1"},{"comment":"Several experimental details are missing, including the batch size, the number of iterations per epoch, the exact U-Net configuration (e.g., number of layers and channels at each depth), and any preprocessing steps beyond rotation augmentation; these details are needed for reproducibility.","section":"Section 2.2"},{"comment":"The Softmax row uses two trainable parameters while the Clip and Sigmoid rows use one, so the comparison is not parameter-matched; please discuss whether the extra parameter affects the comparison.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the contribution is incremental—replacing Clip with Sigmoid and adding GN16 after an IN+LN combination—and the central weakness is the experimental protocol. Even with added seeds and significance tests, the effect sizes are small (around 1-3% DSC) and the evaluation is limited to a single architecture and two datasets. The paper reads more like a workshop submission than a journal paper in its current form. Also, the authors rely on their own prior work [16] to justify excluding BN and GN16 baselines from the main table; the editor may wish to verify that [16] actually contains those comparisons."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a clearly written, modest empirical paper that proposes ILN, a normalization layer which mixes Instance Normalization and Layer Normalization via a sigmoid-weighted sum and then applies Group Normalization with 16 groups. The specific design is new and the motivation is sensible: it fixes two known issues in Batch-Instance Normalization (non-differentiable clip, and loss of normalized distribution after combining). The authors also take care to re-implement all baselines in the same low-level TensorFlow framework and to report overhead. Credit where due: the method is simple, plausible, and the paper is honest about the one validation where it does not help.\n\nThe problem is the evidence. Table 3 shows ILN ahead of IN, LN, GN4 and no-normalization by 0.5–2.5 Dice points, but the per-subject standard deviations are 0.098–0.323, an order of magnitude larger. Each cell is a single run with only two training epochs, and the learning rate was chosen per method after seeing results. That protocol makes the reported numbers upper bounds and the ranking could easily flip with a different seed or split. There are no repeated runs and no significance tests. So the headline claim of 'noticeable accuracy improvements' is not supported by the data. I also note that the original BIN baseline is never tested in the main comparison; Table 1 compares clip vs sigmoid vs softmax for combining IN/LN, which is not the same as BIN's BN/IN combination. No code or data are released, which hurts reproducibility.\n\nThe paper itself is not a bad paper. The idea is reasonable, the write-up is clear, and the authors are candid about limitations (they explicitly acknowledge the LV-3 failure and say the ground-truth rho is unknown). The weaknesses are in the evaluation, not in the reasoning. With repeated seeds, longer training, significance tests, and ideally a released implementation, the claims could be tested properly. As it stands, I'd treat the 1–2% Dice improvement as unverified.\n\nFor peer review: I'd send this to a serious referee. The topic is relevant, the method is a reasonable incremental contribution, and the shortcomings are fixable. But I would not cite it in my own work yet, and I wouldn't bring it to reading group as a strong result—maybe as an example of how easy it is to over-claim from underpowered experiments.","headline":"Plausible normalization variant, but the ~2% Dice claim rests on single two-epoch runs with learning-rate selection after the fact; the evidence is too weak to support it.","tokens_in":7019,"tokens_out":2764,"would_cite":false,"duration_ms":26363,"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":"This paper argues that a trainable blend of Instance Normalization and Layer Normalization, followed by Group Normalization, gives a U-Net a consistent accuracy edge in cardiac MRI segmentation, with reported Dice gains of almost 2% over…","keywords":["Instance-Layer Normalization","U-Net","cardiac MRI segmentation","right ventricle segmentation","left ventricle segmentation","group normalization","Dice similarity coefficient","biomedical image segmentation"],"falsifier":"Re-run the same U-Net and the same RV/LV folds with each normalization method (None, IN, LN, GN4, ILN) under many random initializations and a fixed learning-rate schedule, then compare Dice distributions. If ILN's advantage over the best baseline falls within the inter-run spread, the central claim of consistent improvement is not supported.","tokens_in":5986,"feed_emoji":"🫀","tokens_out":7894,"duration_ms":71696,"temperature":0.7,"pith_summary":"Every convolutional neural network used for medical image segmentation inserts normalization layers that rescale intermediate feature maps for stable training. This paper argues that the normalization strategy need not be fixed per layer, and proposes Instance-Layer Normalization (ILN): normalize the same feature map with instance statistics and with layer statistics, blend the two normalized versions with a trainable weight passed through a sigmoid, and then apply group normalization with 16 groups to restore a normalized distribution. On U-Net segmentation of the right and left ventricles in cardiac MRI, the paper reports that ILN achieves higher Dice scores than no normalization, Instance Normalization, Layer Normalization, and Group Normalization in five of the six cross-validation folds, with gains described as almost 2% Dice. The authors present ILN as a simple drop-in layer whose two design choices—a differentiable blend and a renormalization cascade—fix known problems in Batch-Instance Normalization.","feed_headline":"Sigmoid blend of two normalizers lifts cardiac MRI segmentation by ~2%","feed_subtitle":"A trainable sigmoid blend of instance and layer statistics, renormalized, beats U-Net baselines in five of six folds.","key_machinery":"The load-bearing object is the ILN layer itself, a composition of three standard operations: a per-sample instance normalization over $(H,W)$, a per-sample layer normalization over $(H,W,C)$, and a group normalization with 16 groups applied after their convex combination. The blend weight is $\\rho$ transformed by the sigmoid, so the layer can interpolate continuously between instance and layer statistics and the gradient never hits a non-differentiable corner. The added GN16 stage is what restores zero-mean, unit-variance statistics to the blended feature map, which the paper argues is necessary for signal propagation in deep networks.","core_discovery":"The paper's central discovery is that combining Instance Normalization and Layer Normalization with a smooth, trainable weight, and then renormalizing the combination, is better than using any one of the standard normalization schemes alone. Formally, for a feature map $F$ of shape $(N,H,W,C)$, ILN computes instance-normalized $\\hat{F}^I$ and layer-normalized $\\hat{F}^L$, forms $\\hat{F}^{IL}=\\sigma(\\rho)\\hat{F}^I+(1-\\sigma(\\rho))\\hat{F}^L$ where $\\sigma$ is the sigmoid and $\\rho$ is a trainable scalar, then applies Group Normalization with 16 channel groups to $\\hat{F}^{IL}$ and finishes with per-channel affine parameters $\\gamma,\\beta$. The authors claim this fixes two defects of Batch-Instance Normalization: the Clip function's non-differentiability at 0 and 1, and the loss of a normalized distribution after blending. Their experiments on 6082 right-ventricle images and 805 left-ventricle images with a U-Net show ILN ahead of the baselines in most folds, which they summarize as noticeable accuracy improvements of almost 2% DSC.","pith_inferences":["I would not treat the claimed ranking as stable until multi-seed runs confirm it: Table 3's standard deviations (about 0.10–0.32 Dice) are several times larger than the mean differences between methods (0.005–0.025), and each method was run once for two epochs per fold.","A natural extension is to apply the same sigmoid-blend-plus-renormalize recipe to other normalization pairs, such as Batch plus Instance or Group plus Layer, and to other network families; the paper frames ILN as generic but validates it only on U-Net segmentation.","The reported advantage may be partly an optimization effect—the smooth blend and renormalization may change the effective learning-rate landscape—rather than a property of the normalized statistics themselves; fixing learning-rate selection across methods would separate these."],"forward_implications":["A correct ILN would give medical-image segmentation a drop-in normalization layer: no new loss term, no architectural change, and only one extra trainable scalar per layer plus a group-normalization pass.","On cardiac MRI, gains of almost 2% Dice would translate into better ventricle boundary overlap in a clinical imaging pipeline at negligible additional cost.","Because the blend is differentiable everywhere, gradient-based training can smoothly adapt the normalization regime layer by layer, which the paper shows through layer-specific $\\rho$ training curves.","The GN16 cascade suggests a general recipe: whenever a normalization method produces a combined feature map that is no longer normalized, a subsequent normalization stage can repair the distribution."],"supporting_citations":[{"why":"Supplies the Layer Normalization statistics used as one branch of the ILN combination.","marker":"[1]"},{"why":"Defines the Batch-Instance Normalization baseline whose Clip non-differentiability and unnormalized blend motivate ILN.","marker":"[5]"},{"why":"Provides the SunnyBrook left-ventricle MRI dataset used in the LV segmentation experiments.","marker":"[6]"},{"why":"Supplies the U-Net architecture on which all normalization methods are compared.","marker":"[7]"},{"why":"Supplies the Instance Normalization statistics used as the other branch of the ILN combination.","marker":"[10]"},{"why":"Supplies the Group Normalization operation, specifically GN16, used to renormalize the blended feature map.","marker":"[12]"},{"why":"Provides the prior U-Net normalization comparison and the experimental baseline that ILN extends.","marker":"[16]"},{"why":"Provides the right-ventricle MRI dataset, rotations, and ground truth used in the RV segmentation experiments.","marker":"[17]"},{"why":"Sets the two-epoch and learning-rate protocol followed by all experiments.","marker":"[18]"}],"fun_headline_variants":["Sigmoid blend of instance and layer norms lifts MRI Dice","ILN: trainable norm mix improves cardiac segmentation by 2%","Instance-Layer Normalization: sigmoid combo beats baselines","Sigmoid in norm blend, then group renorm, wins in U-Net"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The largest assumption is that the 0.005–0.025 Dice differences between normalization methods are real method effects rather than training noise, because each method was evaluated with a single two-epoch run per fold and the reported standard deviations are several times larger than those differences.","fun_headline_variants_meta":{"raw":{"variants":["Sigmoid blend of instance and layer norms lifts MRI Dice","ILN: trainable norm mix improves cardiac segmentation by 2%","Instance-Layer Normalization: sigmoid combo beats baselines","Sigmoid in norm blend, then group renorm, wins in U-Net"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0003,"raw_usage":{"total_tokens":1765,"prompt_tokens":1011,"completion_tokens":754,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":676}},"tokens_in":627,"tokens_out":754,"duration_ms":8143,"temperature":1.0,"reasoning_tokens":676,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:55:13.392053+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same U-Net and the same RV/LV folds with each normalization method (None, IN, LN, GN4, ILN) under many random initializations and a fixed learning-rate schedule, then compare Dice distributions. If ILN's advantage over the best baseline falls within the inter-run spread, the central claim of consistent improvement is not supported.","supporting_citations":[{"cited_title":"Stat 1050, 21 (2016)","cited_arxiv_id":null,"evidence_quote":"Supplies the Layer Normalization statistics used as one branch of the ILN combination."},{"cited_title":"In: NeurIPS","cited_arxiv_id":null,"evidence_quote":"Defines the Batch-Instance Normalization baseline whose Clip non-differentiability and unnormalized blend motivate ILN."},{"cited_title":"The MIDAS Journal- Cardiac MR Left Ventricle Segmentation Challenge 49 (2009)","cited_arxiv_id":null,"evidence_quote":"Provides the SunnyBrook left-ventricle MRI dataset used in the LV segmentation experiments."},{"cited_title":"In: ECCV","cited_arxiv_id":null,"evidence_quote":"Supplies the Group Normalization operation, specifically GN16, used to renormalize the blended feature map."},{"cited_title":"IEEE RAL (2019)","cited_arxiv_id":null,"evidence_quote":"Provides the prior U-Net normalization comparison and the experimental baseline that ILN extends."},{"cited_title":"MedIA 44, 86–97 (2018)","cited_arxiv_id":null,"evidence_quote":"Provides the right-ventricle MRI dataset, rotations, and ground truth used in the RV segmentation experiments."}],"review_version":1}