{"id":"09128d83-874b-4f94-92b2-debbf7f579cc","arxiv_id":"2506.18335","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"MCADS, a decoder with depth-to-space upsampling and residual linear attention, improves biomarker segmentation by about three to four percent IoU over previous methods on four public datasets.","lead":"This paper presents MCADS, a new decoder design for medical image segmentation that uses depth-to-space upsampling, channel and spatial attention, and residual linear attention to fuse multiscale features from an encoder. The method reports IoU gains of 2.76 to 4.03 percentage points over prior state-of-the-art on four public nucleus and mitochondria datasets, with code on GitHub.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline tuning and TNBC zero-shot protocol are under-specified, leaving the 2.76-4.03 pt SOTA margins unsupported; no std dev or significance tests are given.","rationale":"The reader's CONDITIONAL verdict matches my read. The decoder modules are described with enough specificity to be reproducible, and the ablations show internal consistency (Tables 5-8), which is real evidence. The weakness is in the external comparison: no variance measures, a protocol asymmetry in patch sizes, and an under-specified zero-shot TNBC table make the headline margins unverifiable from the paper alone. Running standardized re-evaluations would settle the claim. I therefore keep CONDITIONAL/UNCHANGED rather than moving to accept or reject.","tokens_in":15466,"tokens_out":4801,"duration_ms":48531,"concrete_test":"Using the released code, re-run EMCAD and PVT-CASCADE on MoNuSeg and TNBC under an identical protocol: same 256x256 patch size/stride, same augmentation schedule, same epoch count and optimizer, same inference stitching, and at least five seeds; report mean±std per method. Also re-run the TNBC zero-shot evaluation with all models trained on exactly the same source split (e.g., MoNuSeg train) and without any TNBC-fine-tuned entry in the comparison. If EMCAD's MoNuSeg IoU rises above ~72 or the TNBC margin over EMCAD drops below ~1 point, the claimed SOTA advantage is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is comparative: MCADS beats published SOTA by 2.76-4.03 IoU. The load-bearing assumption is that the reproduced baselines are fair, comparably tuned, and that run-to-run spread is smaller than the reported margins. Section 4.2 says all results were reproduced and averaged over five runs, but Tables 1-4 show only point estimates, no standard deviations or significance tests, and several entries are marked '*' as copied from original papers, directly contradicting 'we reproduced all results.' Additionally, Section 4.2 uses 256x256 patches for the proposed method but 224x224 patches for methods with pre-trained backbones; this protocol asymmetry can shift IoU independently of decoder quality. The TNBC gain of 4.03% is the most fragile: Table 4 is described as zero-shot, but no source training split for the zero-shot models is given, and UN-SAM (72.27, above Ours at 69.16) is included despite being fine-tuned on TNBC, so it is not a zero-shot comparison. If baseline tuning were equalized or variances reported, the 2-4 point margins could shrink below significance.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MCADS, a new decoder for medical image segmentation that combines depth-to-space upsampling (DSUB), an effective upsampling block (EUB), channel and spatial attention (CASAB), and a residual linear attention block (RLAB), together with a modified U2-Net encoder that incorporates inner attention and CASAB. The method is evaluated on four datasets (MoNuSeg, DSB, EM, TNBC) against CNN-, transformer-, and SAM-based baselines, and the authors report absolute IoU gains of 2.76–4.03 points over prior state-of-the-art methods. Ablation studies on MoNuSeg examine the contribution of each architectural component and the placement of DSUB stages. The paper also reports inference-time and complexity comparisons and provides a public code link.","tokens_in":15665,"tokens_out":2846,"duration_ms":30056,"significance":"If the reported gains are statistically robust, the paper makes a useful empirical contribution: it shows that a carefully designed decoder can yield competitive or better segmentation accuracy than methods relying on pre-trained backbones, and it provides a systematic ablation of the decoder components. Strengths include the breadth of datasets, the reproduced baselines (for most entries), the ablation coverage, and the public code release. However, the central claim of outperforming state-of-the-art methods rests on five-run averages without variance estimates or significance tests, and the zero-shot comparison on TNBC is not cleanly specified. These issues currently weaken the evidentiary value of the headline margins and need to be addressed before the comparative claims can be fully assessed.","major_comments":[{"comment":"The central claim of outperforming SOTA by 2.76–4.03 IoU points rests on five-run averages reported as point estimates without standard deviations, confidence intervals, or significance tests. Since the reported margins are only a few points and the baseline spread is unknown, the claim is not yet supported. Please report variance across runs and perform paired significance tests (e.g., paired t-test or Wilcoxon signed-rank test over runs, or bootstrap over test images) against the closest baselines such as EMCAD and PVT-CASCADE.","section":"Section 4.2, Tables 1–4"},{"comment":"The TNBC 'zero-shot' comparison is under-specified and partly invalid: the paper does not state which training split the zero-shot models were trained on, and it includes UN-SAM, which the text acknowledges was fine-tuned on TNBC, in the same table without a distinguishing marker. Because UN-SAM (72.27 IoU) is not a zero-shot result, its inclusion makes the comparative table misleading. Please clarify the exact training set used for all zero-shot models, and either remove UN-SAM from the zero-shot table or clearly mark it as a fine-tuned reference.","section":"Section 4.2, Table 4"},{"comment":"The statement 'we reproduced all results' in Section 4.2 and the same claim in the table captions is contradicted by the asterisk-marked entries (InstaSAM, All-in-SAM, UN-SAM, SAC 0-expert in Tables 1, 2, and 4) which are copied from original papers rather than reproduced. This discrepancy undermines the fairness of the comparisons because the copied entries may follow different evaluation protocols. Please state exactly which entries were reproduced and which were taken from the literature, and justify that the evaluation protocols are otherwise comparable.","section":"Section 4.2, Tables 1–4"},{"comment":"The experimental setup uses 256x256 training patches for the proposed method and end-to-end baselines, but 224x224 patches for methods with pre-trained backbones. Patch size can affect segmentation IoU independently of decoder quality, so this asymmetry introduces a confound in the comparisons against pre-trained-backbone methods (e.g., PVT-CASCADE, EMCAD, UCTransNet). Please either use a common patch size for all methods or provide an ablation showing that the reported margins are insensitive to patch size (for example, report the proposed method at 224x224 and at least one major baseline at 256x256).","section":"Section 4.2, patch-size protocol"},{"comment":"The ablation study chooses DSUB in the first two decoder stages as the final configuration, but the table shows that using DSUB in more stages continues to improve IoU (from 74.04 to 74.88 when DSUB is used in all five stages). The selection is justified by a performance/complexity tradeoff, which is a legitimate design choice, but the paper should state explicitly that the selected configuration is not the accuracy-optimal one and should report the sensitivity of the main SOTA comparison to this choice, since a different point on the tradeoff could affect the reported margins.","section":"Table 6 and Section 5.2"}],"minor_comments":[{"comment":"The text says the number of filters F is the product of output channels and the square of the downsampling factor, but writes 'F=C prev×2 d'; please disambiguate the notation (e.g., F = C_prev × 2^d) and clarify which 'output channels' are meant.","section":"Section 3.1.2, Equation (2)"},{"comment":"Equation (8) omits the iteration index i that appears in the surrounding text; please make the notation consistent so each residual iteration is clearly indexed.","section":"Section 3.1.5, Equation (8)"},{"comment":"'M repents the encoder stage' appears to be a typo for 'M represents the encoder stage'.","section":"Section 3.2, Figure 2"},{"comment":"'Coummunications' is misspelled; it should be 'Communications'.","section":"Acknowledgments"},{"comment":"The inference-time comparison reports only EMCAD and the proposed method; please clarify whether the comparison is on the same hardware and whether the times are averaged over the same number of runs, and consider reporting the variance of these timing measurements.","section":"Section 4.3, Computational Cost"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is clearly a CVPR 2025 paper submitted to a journal; the reviewer should be aware that the statistical rigor expected for a journal version is higher than for a conference paper. The core architectural contribution appears sound, but the comparative evidence needs the requested variance, protocol clarification, and TNBC cleanup. If the authors provide those, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a competent empirical decoder design, but the headline gains are not as solid as the abstract suggests. What is new is the specific combination of depth-to-space upsampling in the early decoder stages with a residual linear attention block and a channel-spatial attention block. Each component has clear precedents, and the authors are upfront about the lineage. The ablation study is the strongest part: they systematically remove/add DSUB, EUB, RLAB, CASAB and show progressive IoU gains on MoNuSeg, and they test the decoder across multiple encoders. Code is available. That is a useful engineering contribution.\n\nThe soft spots are real, and the stress-test note is on target. The main claim is comparative: 2.76-4.03 IoU points over SOTA. The tables report five-run averages with no standard deviations or significance tests. We cannot tell whether a 2-3 point margin is larger than run-to-run spread. The patch-size asymmetry is a direct confound: 256x256 for the proposed method, 224x224 for pre-trained backbones. That can shift IoU independently of decoder quality. The TNBC zero-shot protocol is not specified: no training split is given, and UN-SAM is included even though it was fine-tuned on TNBC. The authors acknowledge this in the text, but the abstract's '4.03% over SOTA' is still misleading because UN-SAM outperforms them on that table. Finally, the claim 'we reproduced all results' contradicts the table footnotes that some results were copied from the original papers.\n\nThese concerns do not sink the paper; the decoder may well be better than EMCAD. But they make the current evidence conditional. The central argument holds up as a proof-of-concept, not as a definitive SOTA claim.\n\nI would send this to peer review: it is a serious empirical paper with code, ablations, and reproduced baselines. The authors should add variance estimates, equalize the patch protocol, and specify the zero-shot setup before the 'outperforms SOTA' phrasing is justified. For a reading group, it is a good case study in how to (and how not to) make comparative claims in medical image segmentation.","headline":"Solid empirical decoder paper whose SOTA claims are undercut by missing error bars and a patch-size asymmetry; worth reviewing, but needs revision.","tokens_in":16225,"tokens_out":3599,"would_cite":true,"duration_ms":35021,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A decoder redesign lifts biomarker segmentation by up to 4 percent IoU, without pre-trained encoders.","keywords":["MCADS decoder","medical image segmentation","biomarker segmentation","depth-to-space upsampling","residual linear attention","channel and spatial attention","U2-Net","nuclei segmentation"],"falsifier":"Run MCADS and the strongest baseline (EMCAD) on MoNuSeg for ten independent seeds and compare the per-seed IoU distributions; if the mean difference is within one pooled standard deviation, the claimed state-of-the-art gain is not statistically separable from noise. Alternatively, an independent reproduction using the released code with the paper's exact preprocessing and patch protocol should reproduce the 2.76% margin.","tokens_in":15239,"feed_emoji":"🔬","tokens_out":4781,"duration_ms":41622,"temperature":0.7,"pith_summary":"The paper argues that the decoder, not the encoder, is the bottleneck in medical image segmentation, and that a decoder built from depth-to-space upsampling, channel-spatial attention, and residual linear attention can extract more from the same encoder features. Its proposed MCADS decoder, trained end-to-end from scratch, is claimed to beat published state-of-the-art models—including ones that rely on pre-trained backbones—by absolute IoU margins of 2.76% on MoNuSeg, 3.12% on DSB, 2.87% on Electron Microscopy, and 4.03% on TNBC. The paper also integrates attention into a U2-Net-style encoder and shows the decoder transfers to ResNet-50, MaxViT-B, and PVT encoders. A sympathetic reader would take this as evidence that careful decoder design can close the gap with pre-trained methods.","feed_headline":"New decoder lifts biomarker segmentation by up to 4 percent IoU","feed_subtitle":"Trained from scratch, the MCADS decoder outperforms pre-trained models on MoNuSeg, DSB, EM, and TNBC.","key_machinery":"The load-bearing mechanism is depth-to-space upsampling combined with residual linear attention. DSUB computes the filter count needed for a depth-to-space (sub-pixel) rearrangement—filters equal output channels times the square of the downsampling factor—so that the feature map is reshaped from channel-heavy to spatially detailed without content loss. RLAB iteratively refines each skip connection with a residual block, concatenates it with the decoder feature map, projects to queries, keys, and values, applies scaled dot-product attention, and adds the attended output back through a residual connection. CASAB completes the pathway by combining channel attention (global average and max pooling) with spatial attention built from mean, max, min, and sum pooling.","core_discovery":"The central claim is that three decoder components—Depth-to-Space Upsampling Blocks (DSUB) applied in the first two stages, an Effective Upsampling Block (EUB) in later stages, a Channel and Spatial Attention Block (CASAB), and a Residual Linear Attention Block (RLAB) on skip connections—jointly produce more accurate biomarker segmentation than existing decoders. On the paper's own terms, the depth-to-space operation preserves encoder detail that nearest-neighbor upsampling loses, while RLAB lets the decoder integrate refined encoder features without discarding its own, and CASAB selects the channels and spatial regions that matter. The result is a decoder that is encoder-agnostic and, when paired with a modified U2-Net encoder, achieves the reported gains.","pith_inferences":["A direct test of the paper's encoder-limitation argument would be a scaling curve: holding the MCADS decoder fixed while scaling encoder width and depth, to see where accuracy plateaus.","The same decoder recipe could be applied to other dense prediction tasks (instance masks, depth, surface normals) where high-frequency detail is lost during upsampling.","Because no variance is reported, the central numerical claim is falsifiable by re-running with many seeds; the published margins, not the architecture, are the quantity to scrutinize."],"forward_implications":["End-to-end training from scratch can match or beat pre-trained-encoder methods on four medical segmentation benchmarks, so pre-training is not a prerequisite for state-of-the-art accuracy.","The decoder's gains transfer across CNN and Transformer encoders (ResNet-50, MaxViT-B, PVT v2-B2, modified U2-Net), so it can be dropped into existing architectures.","The ablation studies indicate the first two decoder stages carry most of the upsampling benefit; later DSUB stages add under one IoU point while roughly tripling GFlops.","Boundary metrics (HD95, ASD) improve by larger relative amounts than IoU, implying fewer fragmented or over-segmented regions."],"supporting_citations":[{"why":"EMCAD is the primary state-of-the-art decoder baseline that MCADS is compared against and whose design choices (nearest-neighbor upsampling, attention gate) are contrasted.","marker":"[41]"},{"why":"U2-Net supplies the residual U-block encoder that the paper modifies and serves as the baseline encoder in the ablation study of Table 5.","marker":"[38]"},{"why":"U-Net is the foundational encoder-decoder architecture and is reproduced as a baseline on all four datasets.","marker":"[44]"},{"why":"PVT-CASCADE is a transformer-based cascaded attention decoder baseline reproduced for comparison.","marker":"[40]"},{"why":"UCTransNet is a transformer-based skip-connection method reproduced as a baseline.","marker":"[48]"},{"why":"This work supplies the depth-to-space (sub-pixel) convolution operation used in DSUB.","marker":"[45]"}],"fun_headline_variants":["Decoder without pretraining lifts biomarker segmentation by up to 4% IoU","MCADS decoder boosts biomarker segmentation up to 4.03% IoU","Depth-to-space decoder design beats pretrained models on biomarkers","Trained from scratch decoder outperforms SOTA on four biomarker datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes the reproduced baseline scores, averaged over five runs, reflect each method's best achievable performance under the same tuning protocol used for MCADS, and that the reported 2-4 point IoU margins exceed run-to-run variability; no standard deviations or significance tests are reported (Section 4.2, Tables 1-4).","fun_headline_variants_meta":{"raw":{"variants":["Decoder without pretraining lifts biomarker segmentation by up to 4% IoU","MCADS decoder boosts biomarker segmentation up to 4.03% IoU","Depth-to-space decoder design beats pretrained models on biomarkers","Trained from scratch decoder outperforms SOTA on four biomarker datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000756,"raw_usage":{"total_tokens":3354,"prompt_tokens":935,"completion_tokens":2419,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":2342}},"tokens_in":551,"tokens_out":2419,"duration_ms":17918,"temperature":1.0,"reasoning_tokens":2342,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:50:48.009100+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run MCADS and the strongest baseline (EMCAD) on MoNuSeg for ten independent seeds and compare the per-seed IoU distributions; if the mean difference is within one pooled standard deviation, the claimed state-of-the-art gain is not statistically separable from noise. Alternatively, an independent reproduction using the released code with the paper's exact preprocessing and patch protocol should reproduce the 2.76% margin.","supporting_citations":[{"cited_title":"Emcad: Efficient multi-scale convolutional atten- tion decoding for medical image segmentation","cited_arxiv_id":null,"evidence_quote":"EMCAD is the primary state-of-the-art decoder baseline that MCADS is compared against and whose design choices (nearest-neighbor upsampling, attention gate) are contrasted."},{"cited_title":"U2-net: Go- ing deeper with nested u-structure for salient object detec- tion.Pattern recognition, 106:107404, 2020","cited_arxiv_id":null,"evidence_quote":"U2-Net supplies the residual U-block encoder that the paper modifies and serves as the baseline encoder in the ablation study of Table 5."},{"cited_title":"U- net: Convolutional networks for biomedical image segmen- tation","cited_arxiv_id":null,"evidence_quote":"U-Net is the foundational encoder-decoder architecture and is reproduced as a baseline on all four datasets."},{"cited_title":"Medical im- age segmentation via cascaded attention decoding","cited_arxiv_id":null,"evidence_quote":"PVT-CASCADE is a transformer-based cascaded attention decoder baseline reproduced for comparison."},{"cited_title":"Uctransnet: rethinking the skip connections in u-net from a channel-wise perspective with transformer","cited_arxiv_id":null,"evidence_quote":"UCTransNet is a transformer-based skip-connection method reproduced as a baseline."}],"review_version":2}