{"id":"079b129d-1c46-4d10-aa17-b3ef7eac1569","arxiv_id":"2507.15524","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"RARE-UNet reports the highest average Dice across 1x to 1/8x resolutions on hippocampus and brain tumor MRI segmentation, with faster low-resolution inference.","lead":"This paper describes RARE-UNet, a U-Net variant that sends low-resolution MRI scans directly into deeper layers of a shared encoder, avoiding resampling and cutting compute. On two brain MRI benchmarks it reports the best average Dice across resolutions, but the gain over a simpler augmentation baseline is small and some evaluation details are unclear.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Low-resolution Dice for RARE-UNet may be computed on native coarse grids while baselines are scored at full resolution; if so, the headline average-Dice advantage is an evaluation artifact.","rationale":"The reader's weakest_assumption matches my own reading: the paper's central claim is an empirical superiority claim, and the evaluation protocol is the load-bearing link. The evidence for the concern is textual: Section 3.3 describes resolution-specific heads trained against downsampled ground truth; Section 4.2 specifies input upsampling or padding only for the UNet baselines. No sentence states that RARE-UNet's test-time predictions are resampled to the original resolution before DSC. Because the reported margins over the strongest baseline are tiny on hippocampus (0.838 vs 0.835) and modest on tumor (0.651 vs 0.624), the outcome of the proposed rescoring determines whether the paper's main claim survives. I do not treat this as evidence of misconduct; an underspecified protocol is the most charitable and most probable explanation, and the authors' promised code can settle it. Secondary weaknesses (no significance tests, qualitative inference-time claims, hyperparameter sweep described only for RARE-UNet) reinforce the conditional verdict but are not the single decisive flaw. Verdict recommendation: UNCHANGED relative to the reader's CONDITIONAL, because the concern is real, checkable, and already captured by the conditional.","tokens_in":10739,"tokens_out":7964,"duration_ms":85535,"concrete_test":"Using the released evaluation code from https://github.com/simonsejse/RARE-UNet, inspect the metric computation for RARE-UNet at scales 1/2, 1/4, and 1/8. The decisive check: recompute Tables 1 and 2 by (a) trilinearly upsampling RARE-UNet's low-resolution predictions to the original voxel size and computing DSC against full-resolution ground truth (the protocol implied for UNet-Up baselines), and (b) computing the same downsampled-GT DSC at native scale for all baselines. If (a) reduces RARE-UNet's average margin over UNet+Aug-Up to below about 0.01, or if (b) raises UNet+Aug-Up's low-resolution DSC to match RARE-UNet's, the headline robustness advantage is a metric-resolution artifact and the claim needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing empirical claim is the abstract's 'highest average Dice of 0.84 and 0.65 across resolution.' That claim is only valid if all models are scored against the same ground-truth resolution. The paper never states that RARE-UNet's low-resolution outputs are upsampled before computing Dice. Section 4.2 says only 'For UNet-based models, low-resolution inputs are either padded (Pad) or upsampled (Up) to match the original resolution,' while Section 3.3 explains that each resolution path is trained against a downsampled version of the ground truth via nearest-neighbor interpolation. If RARE-UNet's scale-1/2, -1/4, and -1/8 Dice are computed at native output scale against downsampled GT, while UNet+Aug-Up and nnUNet are scored at full resolution after upsampling or padding, the comparison is not apples-to-apples: baselines must reconstruct fine structure from degraded inputs, whereas RARE-UNet only needs to segment the coarse target. The margins are small enough that this matters: on hippocampus, RARE-UNet's overall 0.838 is only 0.003 above UNet+Aug-Up's 0.835, and at scale 1/8 RARE-UNet (0.785) is actually below UNet+Aug-Up (0.789); on tumor the average margin is 0.027. Rescoring RARE-UNet at full resolution could easily erase or reverse these differences. No significance test is reported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RARE-UNet, a 3D UNet extension with multi-scale gateway blocks (MSBs) placed at different encoder depths. Inputs downsampled by factors of 2^d are routed to the MSB at depth d, bypassing shallower layers, and each resolution path has its own segmentation head. Training combines a per-path segmentation loss (cross-entropy and soft Dice) with an MSE consistency loss between MSB features and the corresponding full-resolution encoder features. Experiments compare against plain UNet, a multi-resolution-augmented UNet (UNet+Aug), and nnUNet at full, 1/2, 1/4, and 1/8 resolutions on two Medical Segmentation Decathlon tasks (hippocampus and brain tumor), reporting Dice scores and inference times. The paper claims the highest average Dice across resolutions (0.84 and 0.65) and significantly reduced inference time at low resolutions.","tokens_in":11022,"tokens_out":5859,"duration_ms":64626,"significance":"The architecture is well motivated and the code is publicly released. If the empirical claims hold under a uniform evaluation protocol, RARE-UNet would be a practical single-model solution for variable-resolution medical-image segmentation, with plausible computational savings at low resolutions. The comparison includes strong baselines, including nnUNet and a dedicated multi-resolution augmentation variant, and the consistency-loss mechanism is a reasonable way to align features across scales. The main risk is that the headline quantitative claim depends on an ambiguous metric-resolution protocol and on mean differences that are not tested for significance; both issues are fixable within the scope of a revision.","major_comments":[{"comment":"The evaluation protocol for computing Dice at reduced scales is not specified for RARE-UNet. Section 4.2 states that for UNet-based models low-resolution inputs are either padded or upsampled to the original resolution, so baselines are scored against full-resolution ground truth. Section 3.3, however, says each RARE-UNet resolution path is trained against a downsampled version of the ground truth using nearest-neighbor interpolation. The text never states whether RARE-UNet's low-resolution outputs are upsampled to full resolution before computing Dice. If they are scored at native coarse resolution against downsampled ground truth, the abstract's 'highest average Dice across resolution' is confounded: coarse-grid segmentation does not need to resolve fine boundaries, whereas the upsampled baselines are penalized for boundary inaccuracy. This is not a hypothetical concern given the margins in Tables 1-2 (hippocampus overall 0.838 vs 0.835 for UNet+Aug-Up, and 0.785 vs 0.789 at scale 1/8). Please state the evaluation resolution explicitly for all models and, if RARE-UNet was evaluated natively, rescore its low-resolution outputs after upsampling to full resolution against full-resolution ground truth.","section":"Section 4.2 / Section 3.3"},{"comment":"No statistical significance testing is reported, so the central claims of 'highest' and 'consistently outperforms' are not supported at the reported margins. On hippocampus, RARE-UNet's overall Dice of 0.838 ± 0.033 is within one standard deviation of UNet+Aug-Up's 0.835 ± 0.032, and RARE-UNet is numerically worse than UNet+Aug-Up at scale 1/8 (0.785 vs 0.789). On tumor, the overall margin (0.651 vs 0.624) is larger but still unreported for variability across test subjects. Please add paired significance tests (e.g., Wilcoxon signed-rank or bootstrap confidence intervals on per-subject Dice) for the overall average, for each reduced scale, and for inference-time differences in Fig. 6.","section":"Section 4.3, Tables 1-2"},{"comment":"The claimed resolution-adaptive behavior is only demonstrated for inputs that are exact dyadic downsamplings of the training resolution, with routing based on the known scale. Real-world resolution variation will not fall exactly on these levels, and the inference-time rule 'rounded to the nearest matching resolution level' (Section 3.2) is never tested for off-grid input sizes or for inputs whose dimensions are not multiples of the routing factors. Since adaptability to heterogeneous resolutions is the paper's central motivation, the evaluation should include at least one non-dyadic or shifted-resolution condition, or the claims should be explicitly delimited to dyadic scales.","section":"Section 3.2, Section 4.2"}],"minor_comments":[{"comment":"The caption says 'S0 through S4' but the plot and the text describe four scales (full, 1/2, 1/4, 1/8); correct the caption or the label count.","section":"Figure 6"},{"comment":"The text says RARE-UNet achieves 'dynamic parameter scaling with input resolution,' but Fig. 6 reports a fixed parameter count for each model; only the active computational path changes, not the parameter count. Please rephrase.","section":"Section 4.3"},{"comment":"The paper states that over 100 runs were used to select the learning rate and consistency loss weight, but it does not report the selected values or the validation criterion in detail; please report the chosen hyperparameters or point to a configuration file.","section":"Section 4.2"},{"comment":"Inference times are reported as averages over 10 runs without hardware details or variance, so the claim of 'significantly reduced inference time' is not quantitatively supported; provide the GPU/CPU specification and error bars or a range.","section":"Model Efficiency, Section 4.3"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the RARE-UNet idea is a legitimate engineering contribution, and the stress-test concern is right - the paper never says whether low-resolution RARE-UNet outputs are upsampled before Dice is computed. That has to be fixed before the headline numbers mean anything.\n\nWhat is new: routing downsampled inputs into a shared UNet encoder at matched depths, giving each path its own segmentation head and adding an MSE consistency loss. Individual pieces are known, but I do not know of prior work combining them this way for variable-resolution segmentation, and it is a sensible response to a real clinical problem. The paper explains the architecture and loss clearly, evaluates on two public benchmarks, and includes a resolution-augmented UNet baseline plus nnUNet. The GitHub link is a plus, though I have not verified what is in it.\n\nThe soft spot is the evaluation. Section 3.3 trains each resolution path against a downsampled ground truth. Section 4.2 explains that only the UNet baselines pad or upsample low-resolution inputs to match the original resolution. There is no corresponding statement for RARE-UNet. If RARE-UNet scale-1/4 and 1/8 Dice are computed at native low resolution against downsampled ground truth, while the Up baselines are scored at full resolution, the comparison rewards RARE-UNet for not having to reconstruct fine structure. On hippocampus the margin over UNet+Aug-Up is tiny anyway - 0.838 vs 0.835 overall, and UNet+Aug-Up is actually better at 1/8 (0.789 vs 0.785) - so this ambiguity can flip the central claim. On tumor the advantage (0.651 vs 0.624 average) is more comfortable, but it still depends on the same protocol.\n\nTwo smaller issues. First, no significance tests or confidence intervals; several differences are within one standard deviation. Second, the inference-time speedup is asserted as significantly reduced but the text gives no numeric table; Figure 6 shows a trade-off plot but no hardware, batch size, or repeat-run details. Also, the 100+ run hyperparameter sweep is described for RARE-UNet, but I do not see whether the baselines received comparable tuning.\n\nThis paper is for people working on deployment-oriented variable-resolution medical segmentation, and for reviewers who care about evaluation protocol pitfalls. It is not a field-reorganizing idea, but the architecture has real potential if the numbers hold up.\n\nRecommendation: send it to peer review. The authors need to state explicitly for every model what resolution the prediction is at when Dice is computed, report significance or intervals, and give actual inference-time numbers. If the protocol is fixed and the results survive, this is a solid conference paper.","headline":"Plausible architecture; the unspecified low-resolution Dice evaluation protocol can flip the main result.","tokens_in":11567,"tokens_out":5545,"would_cite":true,"duration_ms":58409,"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":"RARE-UNet, a resolution-aware UNet with multi-scale gateway blocks, claims stable segmentation accuracy down to 1/8-resolution inputs while running four times faster, outperforming UNet, UNet+Aug, and nnUNet on two brain MRI tasks.","keywords":["resolution-aware segmentation","multi-scale UNet","adaptive inference routing","low-resolution robustness","brain MRI segmentation","consistency loss","Dice score"],"falsifier":"Recompute the scale-$1/8$ Dice for all models after upsampling every prediction to full resolution before comparing with full-resolution ground truth, keeping the training and routing identical; if the $1/8$-gap between RARE-UNet and nnUNet or UNet+Aug narrows to near zero, the resolution-robustness claim fails.","tokens_in":10515,"feed_emoji":"🧠","tokens_out":6060,"duration_ms":57608,"temperature":0.7,"pith_summary":"The paper sets out to show that a single UNet-style architecture can segment brain MRI images robustly when input resolution varies, instead of assuming fixed high-resolution inputs. It claims that RARE-UNet, by routing downsampled images into deeper encoder layers through multi-scale gateway blocks and aligning features with a consistency loss, keeps Dice scores nearly flat from full resolution down to $1/8$ scale. On hippocampus and brain tumor segmentation benchmarks, it reports the highest average Dice across resolutions (0.84 and 0.65), surpassing standard UNet, a multi-resolution-augmented UNet, and nnUNet, while cutting inference time at lower resolutions. If true, this would make clinical segmentation pipelines more reliable on heterogeneous, non-standardized scans without heavy resampling preprocessing.","feed_headline":"Resolution-aware UNet keeps low-res MRI Dice high, 4x faster","feed_subtitle":"RARE-UNet routes 1/8-scale scans into deeper encoder layers, holding Dice while cutting computation.","key_machinery":"The central mechanism is the multi-scale gateway block (MSB), a convolutional block placed at several encoder depths that transforms a downsampled input into a feature map aligned in shape and semantics with the encoder output at that depth. An MSE consistency loss enforces this alignment during training, so the block learns to reproduce what the full encoder would have computed had it processed the image at full resolution. At inference, a lightweight routing step rounds the input size to the nearest resolution level and activates only the relevant path, skipping shallow layers for low-resolution inputs. The resolution-specific segmentation heads then produce predictions without global resampling.","core_discovery":"RARE-UNet is a 3D UNet extended with multi-scale gateway blocks that act as resolution-aware entry points at different encoder depths. A full-resolution input enters at the first encoder layer, while inputs downsampled by factors of $1/2$, $1/4$, and $1/8$ bypass shallow layers and are injected at matching depths, sharing the same bottleneck and decoder. Each resolution path has its own segmentation head, and training combines per-path cross-entropy and soft Dice losses with an MSE consistency loss that pushes each gateway block's features toward the encoder features at that depth. The authors report that this design achieves the highest average Dice scores across resolutions—0.84 for hippocampus and 0.65 for brain tumor—with lower variance than baselines, and a roughly $4\\times$ speedup when input resolution is halved.","pith_inferences":["If the Dice metric is later confirmed to be computed at native resolution for RARE-UNet, the practical gain over the upsampling-augmented baseline may be mostly about efficiency rather than accuracy; the $1/8$-scale Dice of 0.785 (hippocampus) versus 0.789 for UNet+Aug-Up is close.","The routing-by-resolution idea could be extended to anisotropic or variable slice-thickness inputs, which the current $1/2$, $1/4$, $1/8$ isotropic downsampling does not cover.","The consistency loss between gateway and encoder features suggests a possible knowledge-distillation view: low-resolution paths are trained to emulate the full-resolution encoder, which may make the gains transferable to other tasks like detection or registration."],"forward_implications":["Heterogeneous multi-center MRI data could be segmented without per-scan resampling or padding, since the network accepts a range of native input resolutions.","Low-resolution scans, which are cheaper and faster to acquire, would not require a separate low-resolution model or an upsampling pre-processing step.","Inference cost scales with input resolution: halving each spatial dimension yields roughly a $4\\times$ speedup, which could matter in time-critical clinical settings.","The same resolution-routing design could be applied to other encoder-decoder architectures beyond UNet, since the gateway blocks only need a feature-alignment signal."],"supporting_citations":[{"why":"The original UNet architecture that RARE-UNet extends and the primary full-resolution baseline.","marker":"[3]"},{"why":"nnUNet, the state-of-the-art self-configuring segmentation framework that serves as the strongest baseline at full resolution.","marker":"[12]"},{"why":"The BraTS brain tumor dataset and challenge, one of the two evaluation tasks.","marker":"[13,14]"},{"why":"The VUMC hippocampus dataset, the second evaluation task.","marker":"[15]"},{"why":"The Medical Segmentation Decathlon preprocessing and data split protocol that defines the input sizes and tasks used in the experiments.","marker":"[16]"}],"fun_headline_variants":["Low-res MRI segmentation: adaptive UNet hits top Dice, 4x faster","RARE-UNet routes low-res inputs deep to keep Dice high and cut compute","Resolution-adaptive UNet: best Dice across scales, 4x faster at half res","Inject low-res MRI at deeper layers: UNet variant wins Dice, 4x speedup","RARE-UNet: adaptive entry points keep Dice high on low-res brain MRIs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation loads on the assumption that Dice scores are computed the same way for every model at every resolution; the paper does not clearly state whether RARE-UNet's low-resolution predictions are upsampled to full resolution before scoring, while the baselines are, so the headline low-resolution gains could be an artifact of comparing scores at different output resolutions.","fun_headline_variants_meta":{"raw":{"variants":["Low-res MRI segmentation: adaptive UNet hits top Dice, 4x faster","RARE-UNet routes low-res inputs deep to keep Dice high and cut compute","Resolution-adaptive UNet: best Dice across scales, 4x faster at half res","Inject low-res MRI at deeper layers: UNet variant wins Dice, 4x speedup","RARE-UNet: adaptive entry points keep Dice high on low-res brain MRIs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000438,"raw_usage":{"total_tokens":2205,"prompt_tokens":904,"completion_tokens":1301,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":1201}},"tokens_in":520,"tokens_out":1301,"duration_ms":10124,"temperature":1.0,"reasoning_tokens":1201,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:29:54.704592+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the scale-$1/8$ Dice for all models after upsampling every prediction to full resolution before comparing with full-resolution ground truth, keeping the training and routing identical; if the $1/8$-gap between RARE-UNet and nnUNet or UNet+Aug narrows to near zero, the resolution-robustness claim fails.","supporting_citations":[{"cited_title":"In: MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, Springer (2015) 234–241 12 S","cited_arxiv_id":null,"evidence_quote":"The original UNet architecture that RARE-UNet extends and the primary full-resolution baseline."},{"cited_title":"Nature Methods18(2) (2021) 203–211","cited_arxiv_id":null,"evidence_quote":"nnUNet, the state-of-the-art self-configuring segmentation framework that serves as the strongest baseline at full resolution."},{"cited_title":"Nature Communications13(1) (2022) 4128","cited_arxiv_id":null,"evidence_quote":"The Medical Segmentation Decathlon preprocessing and data split protocol that defines the input sizes and tasks used in the experiments."}],"review_version":1}