{"id":"9abc1d95-bf02-4e2e-9dcd-9e42457c40b0","arxiv_id":"2506.13415","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"SimpleUNet is an ultra-lightweight U-Net variant that reports accuracy close to or above much larger models on breast ultrasound, skin lesion, and polyp segmentation datasets.","lead":"The paper presents SimpleUNet, a stripped-down U-Net that shrinks skip connections, fixes channel width, and adds learned channel weights, reaching as few as 16,000 parameters while matching or beating much larger models on some medical datasets. If the results hold up, small clinics could run useful segmentation AI on cheap, low-power hardware.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ISIC test-set-as-validation protocol inflates headline mDice values; the abstract's ISIC dominance claim is not independently evaluated.","rationale":"The reader's conditional verdict is appropriate. The ISIC test-set-as-validation protocol is a real, concrete flaw that directly undermines the abstract's ISIC numbers and the 'consistent dominance' claim. I see no reason to move the verdict from CONDITIONAL: the MBD and KVASIR results appear credible and the ablation study is internally coherent, but the ISIC claim needs an independent validation protocol and ideally verified public code before full acceptance. I also note, without making it the primary objection, that the abstract's 0.67 MB variant is actually SimpleESKNet2+AFF32, an ESKNet-block extension rather than the plain fixed-width SimpleUNet, which further tempers the 'simple is all you need' phrasing. The proposed concrete test directly isolates the leakage concern and would settle whether the ISIC numbers survive an honest evaluation protocol.","tokens_in":20429,"tokens_out":6840,"duration_ms":68891,"concrete_test":"Re-run the ISIC2017 and ISIC2018 experiments with a proper held-out validation split: hold out, say, 20% of the official training images, use it only for best-epoch checkpoint selection, then evaluate all checkpoints once on the untouched official test set. Compare SimpleUNet-16KB, SimpleUNet-0.47MB, SimpleESKNet-0.67MB, LBUNet, UNet, TransUNet, and ESKNet under this protocol. If the reported mDice values drop by more than roughly 0.5–1 point or the ranking changes on either ISIC dataset, the abstract's ISIC claims and the 'consistent dominance' wording must be revised; if they survive, the leakage concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-A states: 'For datasets containing validation datasets, we only validate the performance of the model during training after each epoch and iteratively save the weights that give the best performance on the validation set... Otherwise, the testing set is used as the validation set instead.' ISIC2017 and ISIC2018 have no official validation split (Table II lists only Train and Test). Hence every ISIC number in Table XVI, including the abstract's mDice 84.86%/88.77%, is produced by selecting the best epoch on the exact test images used for the reported score. This is not an independent evaluation; it is a form of test-set leakage and can materially overstate accuracy. Because all compared methods are apparently run under the same protocol, relative rankings may be less affected, but the absolute values and the claim of 'consistent dominance' over published state-of-the-art results on ISIC are not supported. The issue is directly load-bearing: two of the four datasets in the abstract are ISIC, and the claim that SimpleUNet 'outperforms LBUNet and other lightweight benchmarks across multiple public datasets' is contradicted by the paper's own Table XVI on ISIC2018 (LBUNet 87.85 vs SimpleUNet 87.57 mDice), so an independent protocol is needed before accepting the headline.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes SimpleUNet, a lightweight U-Net variant built from three strategies: partial channel selection on skip connections, fixed width across all encoder/decoder stages, and a learnable channel-wise weighting of shortcut and deep features. The authors report parameter counts as low as 16 KB and 0.67 MB for their larger variant, with experiments on a merged breast ultrasound dataset (MBD), two skin-lesion datasets (ISIC2017/2018), and a polyp dataset (Kvasir-SEG). Extensive ablations vary width, kernel size, dilation rate, feature-selection rate, number of blocks, and adaptive fusion, and the framework is extended to ESKNet blocks as SimpleESKNet. The central claims are that the 16 KB model outperforms lightweight benchmarks and that the 0.67 MB variant surpasses U-Net and TransUNet on several datasets.","tokens_in":20690,"tokens_out":10542,"duration_ms":97133,"significance":"The architectural idea is attractive and the empirical scope is substantial: every model is trained five times from scratch, all hyperparameters are swept in ablations, and the released code makes the parameter counts directly verifiable. If the reported ISIC numbers were obtained under an independent protocol, the result would be practically important because it would show that extreme parameter compression does not force a large accuracy penalty. The extendability result (replacing convolution blocks with ESKNet blocks) is a useful demonstration that the design pattern is transferable. The main value of the paper, if corrected, would be as a simple, reproducible baseline for efficient medical image segmentation.","major_comments":[{"comment":"The protocol sentence 'Otherwise, the testing set is used as the validation set instead' applies directly to ISIC2017 and ISIC2018, because Table II lists only Train and Test for those datasets. Best-epoch weights are therefore selected on the exact test images that are later scored, so the reported ISIC mDice/mIoU values in Table XVI (including the abstract's 84.86% and 88.77%) are in-sample estimates rather than independent test-set evaluations. This is a load-bearing threat to the headline 'consistent dominance' claim on two of the four datasets; the authors must add a proper validation split (or a pre-registered selection rule) and re-evaluate, or remove/qualify the ISIC claims.","section":"Section IV-A and Table II"},{"comment":"The abstract's statement that the 16 KB configuration 'outperforms LBUNet and other lightweight benchmarks across multiple public datasets' is contradicted by the paper's own Table XVI on ISIC2018: SimpleUNet16 scores 87.57 mDice and 79.98 mIoU, whereas LBUNet scores 87.85 and 80.69. On ISIC2017 the same 16 KB model (83.47 mDice) is also below UNexts (83.79). The headline should be restricted to MBD and Kvasir-SEG, or the claim should be explicitly qualified as 'on some datasets.'","section":"Abstract and Table XVI"},{"comment":"Several of the headline superiority claims are based on differences well within one standard deviation, e.g., MBD SimpleESKNet32 85.76±0.49 vs TransUNet 85.32±0.48 and Kvasir-SEG SimpleESKNet32 86.46±0.42 vs ESKNet 86.60±0.83. With M=5 independent runs and no paired significance test, these differences are not established. The authors should report paired tests (e.g., Wilcoxon signed-rank or bootstrap over test images) for at least the abstract's comparison set, and avoid 'dominating' language where differences are not significant.","section":"Table XVI"}],"minor_comments":[{"comment":"The text twice refers to the 'Hardmard product'; this should be 'Hadamard product.'","section":"Section III-D"},{"comment":"The caption says '(MBT)' but the dataset is called MBD everywhere else; the acronym should be made consistent.","section":"Table I caption"},{"comment":"References [2] and [3] are duplicate entries for the same Mambasam paper and should be merged.","section":"References"},{"comment":"The parameter-count formulas are stated for the case 'N as one,' while Table III experiments use two stacked convolution layers per block; please reconcile the formulas with the exact architectures used in the experiments or explicitly label them as rough estimates for a simplified U-Net.","section":"Section III-A/III-E"},{"comment":"The abstract gives a GitHub link, but the Data Availability section states that code will be released after acceptance; these statements should be made consistent.","section":"Abstract and Data Availability"},{"comment":"The names 'SimpleUNet1_16' and 'SimpleESKNet1+AFF_16' are not explained in a table footnote; the reader cannot tell which subscript/superscript denotes width and which denotes block count.","section":"Table XVI notation"},{"comment":"Write 'Kvasir-SEG' rather than 'KV ASIR-SEG' for consistency with the dataset's official name.","section":"Throughout"},{"comment":"The entry 'Para < 0.251' for the model without adaptive feature fusion is imprecise; report the exact parameter count.","section":"Table XII"}],"recommendation":"major_revision","confidential_remarks":"The main concern is the ISIC evaluation protocol, because the abstract's ISIC numbers are the primary evidence for the word 'dominance.' I would ask the authors to rerun those experiments with a held-out validation split or to drop the ISIC-based claims entirely. The architecture itself and the MBD/Kvasir ablations are worth publishing after this fix."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the good news: the core architecture is genuinely simple and the ablations are thorough. The paper argues that a fixed-width encoder plus channel-reduced skip connections (a 1x1 conv selecting partial features) plus a learned channel-wise scaling can shrink a U-Net to 0.67M parameters without losing much accuracy on breast ultrasound and polyp segmentation. The parameter counts in Section III are arithmetic derivations, not fitted numbers, and the 5-run means with standard deviations are a step up from typical reporting in this literature. The MBD and KVASIR results look credible; the 16KB variant hitting 82.93% DSC on a merged breast dataset is a real data point.\n\nThe large soft spot is the ISIC evaluation. Section IV-A says that when a dataset has no validation split, the testing set is used as the validation set. ISIC2017 and ISIC2018 have no official validation split, so the best-epoch weights are selected on the exact test images used for the reported mDice. That is test-set leakage. The absolute numbers in the abstract (84.86% and 88.77%) are therefore not independent, and the claim of 'consistent dominance' over published results is unsupported. This is not a minor detail; two of the four datasets in the abstract are ISIC. Also check the model naming: those abstract ISIC numbers come from SimpleESKNet, not the SimpleUNet architecture that the paper presents as the main contribution. And on ISIC2018, the 16KB SimpleUNet is actually below LBUNet (87.57 vs 87.85), so the abstract's 'outperforms LBUNet across multiple datasets' is only true for three of the four.\n\nOther, smaller issues: the code availability statement is contradictory ('codes can be found at github' in the abstract versus 'once accepted' in the data availability section), and the paper's own Table XVI shows several differences within overlapping error bars, so the 'dominance' language is overstrong. The paper does acknowledge the GFLOPs are not always competitive because it uses plain convolutions without early spatial reduction; that is honest.\n\nWho should read it? Anyone working on lightweight medical segmentation who wants a clearly-described baseline and a cautionary example of validation leakage. If the ISIC protocol is fixed or re-reported, the abstract needs revision. I would send this to a serious referee: the architecture and MBD/KVASIR experiments deserve scrutiny, and the ISIC methodology needs to be checked. Not a desk reject, but a conditional acceptance after the ISIC claims are corrected.","headline":"A simple, well-ablated lightweight U-Net with credible breast and polyp results, but the ISIC numbers are undermined by test-set-as-validation and the abstract overstates dominance.","tokens_in":21253,"tokens_out":4800,"would_cite":true,"duration_ms":41140,"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 claims that a U-Net reduced to a fixed width with selected skip connections and adaptive feature fusion stays competitive while shrinking to as little as 16 KB, and at 0.67 MB beats far larger models on a breast-lesion benchmark.","keywords":["medical image segmentation","ultra-lightweight segmentation","U-Net","skip-connection feature selection","fixed-width architecture","adaptive feature fusion","parameter efficiency"],"falsifier":"A re-run of ISIC2017/2018 with a validation split disjoint from the test set would settle the issue: if the best-validation-epoch mDice falls below the reported 84.86%/88.77%, the skin-lesion numbers are inflated by the selection procedure.","tokens_in":20223,"feed_emoji":"🩺","tokens_out":11296,"duration_ms":101649,"temperature":0.7,"pith_summary":"The paper sets out to show that an intentionally simple U-Net can be both extremely small and highly accurate for medical image segmentation. Its central claim is that three plain changes — sending only a selected fraction of the skip-connection channels to the decoder, keeping the channel count fixed at $C$ across all stages, and reweighting fused features with two learned vectors — cut the parameter count by about two orders of magnitude while preserving or improving segmentation overlap (Dice/IoU) on breast, skin, and polyp datasets. If true, model size stops being a reliable proxy for quality in this setting, and memory-limited clinical deployment becomes feasible without a large accuracy penalty. The paper reports a 16 KB configuration that outperforms other lightweight models, and a 0.67 MB configuration whose reported Dice/IoU on a merged breast-lesion dataset exceeds both U-Net and TransUNet.","feed_headline":"16 KB medical segmentation network outdoes rival lightweight models","feed_subtitle":"Fixed-width design with selected skip connections also tops U-Net and TransUNet at 0.67 MB.","key_machinery":"Three mechanisms carry the argument. The feature-selection skip connection is a $1\\times1$ convolution that shrinks shortcut channels to $R\\times C$, so the decoder fuses only a representative subset of features. Fixed-width encoding keeps every stage at $C$ channels, converting U-Net's exponential parameter growth into linear growth. Adaptive feature fusion multiplies the selected shortcut and deep features by learnable vectors $\\alpha$ and $\\beta$ (Hadamard product) before concatenation, adding almost no parameters while letting the network weight channels by importance. Together they yield the closed-form parameter estimate $\\sim 6 C^2 k^2$ for the default settings, which is the quantitative fact that supports the claim that width expansion and full skip concatenation are the main sources of U-Net's bulk.","core_discovery":"The paper's core claim is that the two habits that make U-Net large — stagewise doubling of channel width and concatenating every shortcut feature into the decoder — are not what makes it accurate. In SimpleUNet, each skip connection passes through a $1\\times1$ convolution that keeps only $R\\times C$ channels, the encoder and decoder keep a constant width $C$ instead of widening, and a fusion step multiplies the selected shortcut and deep features by learned channel weights $\\alpha$ and $\\beta$ before concatenation. With feature-selection rate $R=0.5$ and one convolution block per stage, the whole network holds roughly $6 C^2 k^2$ parameters, about 65 times fewer than a standard U-Net. The paper reports that the resulting 16 KB model outperforms LBUNet and other lightweight benchmarks on a merged breast-lesion dataset, and that its best 0.67 MB variant reaches mean DSC/IoU of 85.76%/75.60% there, higher than the reported scores of U-Net and TransUNet, with strong mDice on ISIC2017, ISIC2018, and KVASIR-SEG.","pith_inferences":["Editorial extension: the same fixed-width recipe should carry over to 3D volumetric segmentation, where memory limits are tighter; the paper reports only 2D experiments.","Editorial extension: the ablations imply the model's capacity budget is spent on depth and receptive field instead of width, so the gains likely depend on deeper blocks and dilation, not on width.","Editorial extension: the reported ISIC numbers may be optimistic because the test set is used for best-epoch selection when no validation split exists; re-running with a separate validation split would give an unbiased estimate.","Editorial extension: replacing plain convolutions with grouped depthwise or spatially reduced operations, which the discussion names but does not implement, should lower the computational cost that currently remains higher than some lightweight rivals."],"forward_implications":["The 16 KB SimpleUNet can act as a practical strong baseline for memory-constrained medical imaging, outperforming other lightweight models on the merged breast dataset.","The 0.67 MB SimpleESKNet gives a parameter-versus-accuracy target: about 140 times fewer parameters than TransUNet with equal or better reported DSC/IoU on the breast-lesion benchmark.","U-Net's progressive width doubling and full skip concatenation are not essential to accuracy on these tasks; removing them costs little and saves most of the parameters.","The feature-selection and adaptive-fusion strategies are transferable, since replacing the plain blocks with ESKNet's blocks preserves and extends the gains."],"supporting_citations":[{"why":"Defines the base U-Net architecture whose parameter growth is analyzed and reworked.","marker":"[1]"},{"why":"TransUNet is the heavy baseline it must beat; comparison target on all four datasets.","marker":"[5]"},{"why":"ESKNet's convolution blocks are inserted into the SimpleUNet skeleton to form SimpleESKNet, establishing transferability.","marker":"[6]"},{"why":"LBUNet is the lightweight benchmark with 38 KB parameters that the 16 KB SimpleUNet claims to outperform.","marker":"[12]"},{"why":"UltraVMUNet supplies a recent Mamba-based lightweight comparison point.","marker":"[13]"},{"why":"UNext is the strongest lightweight competitor on the merged breast data, used as the main rival for the 16 KB model.","marker":"[16]"},{"why":"ISIC2017 supplies the skin-lesion test set used for the reported mDice.","marker":"[32]"},{"why":"ISIC2018 supplies the larger skin-lesion test set used for the reported mDice.","marker":"[33]"},{"why":"KVASIR-SEG supplies the polyp segmentation test set used for the reported mDice/mIoU.","marker":"[34]"}],"fun_headline_variants":["16KB U-Net beats larger models with simpler design","Simpler design makes 16KB segmentation model outperform U-Net","Tiny 16KB UNet beats heavyweight models via smart skip connections","Ultra-light 16KB model tops U-Net and TransUNet on medical segmentation","Simplicity wins: 16KB segmentation net beats costly U-Net designs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The skin-lesion results rest on the assumption that using the test set to pick the best training epoch and then scoring on that same test set does not inflate the reported numbers, because the paper states that when no validation split exists, the test set is used as the validation set.","fun_headline_variants_meta":{"raw":{"variants":["16KB U-Net beats larger models with simpler design","Simpler design makes 16KB segmentation model outperform U-Net","Tiny 16KB UNet beats heavyweight models via smart skip connections","Ultra-light 16KB model tops U-Net and TransUNet on medical segmentation","Simplicity wins: 16KB segmentation net beats costly U-Net designs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000824,"raw_usage":{"total_tokens":3654,"prompt_tokens":1049,"completion_tokens":2605,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":665,"completion_tokens_details":{"reasoning_tokens":2518}},"tokens_in":665,"tokens_out":2605,"duration_ms":17648,"temperature":1.0,"reasoning_tokens":2518,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:01:59.277980+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A re-run of ISIC2017/2018 with a validation split disjoint from the test set would settle the issue: if the best-validation-epoch mDice falls below the reported 84.86%/88.77%, the skin-lesion numbers are inflated by the selection procedure.","supporting_citations":[{"cited_title":"Esknet: An enhanced adaptive selection kernel convolution for ultrasound breast tumors segmentation,","cited_arxiv_id":null,"evidence_quote":"ESKNet's convolution blocks are inserted into the SimpleUNet skeleton to form SimpleESKNet, establishing transferability."},{"cited_title":"Lb-unet: A lightweight boundary-assisted unet for skin lesion segmentation,","cited_arxiv_id":null,"evidence_quote":"LBUNet is the lightweight benchmark with 38 KB parameters that the 16 KB SimpleUNet claims to outperform."},{"cited_title":"Unext: Mlp-based rapid medical image segmentation network,","cited_arxiv_id":null,"evidence_quote":"UNext is the strongest lightweight competitor on the merged breast data, used as the main rival for the 16 KB model."},{"cited_title":"Bus-bra: A breast ultrasound dataset for assessing computer-aided diagnosis systems,","cited_arxiv_id":null,"evidence_quote":"ISIC2017 supplies the skin-lesion test set used for the reported mDice."}],"review_version":2}