{"id":"57cb3187-4bef-485e-9c15-8bb4294646dc","arxiv_id":"2511.18493","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A CNN-Transformer U-Net with hierarchical expert routing and a shape-adapting hub reports state-of-the-art colon histopathology segmentation Dice of 95.57% on EBHI, 95.16% on DigestPath, and 94.17% on GlaS.","lead":"SAGE is a dynamic mixture-of-experts framework that lets a hybrid CNN-Transformer segmentation network choose which internal 'experts' to use for each tissue image, reporting top Dice scores on three colon cancer histopathology benchmarks. A smart generalist should read it because input-adaptive routing could reduce compute and improve out-of-distribution robustness in medical image segmentation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Validation-set reporting and split/resolution mismatches make the EBHI and DigestPath SOTA claims unverifiable; abstract numbers also disagree with Table 2.","rationale":"The reader's weakest assumption is exactly where I would put the load: the EBHI and DigestPath numbers are validation numbers, the resolution protocol shifts between the baseline and SOTA tables, and the abstract numbers do not match Table 2. Without held-out test evaluation or at least k-fold cross-validation with fixed protocol, the central 'state-of-the-art' claim cannot be verified from the paper. I do not see a separate more fundamental flaw in the routing formulation; the concern is about the evidence supporting the empirical claim, not about the internal consistency of the SAGE equations. Thus the appropriate disposition remains CONDITIONAL — the method may work, but it needs independent held-out validation, consistent reporting, and ideally code/checkpoint release before the SOTA claim can be accepted. I agree with the reader's assessment that the resolution and validation issues are the key obstacle, so no change in verdict is needed.","tokens_in":16375,"tokens_out":5884,"duration_ms":57637,"concrete_test":"Run SAGE-UNet on the official DigestPath test set (via the challenge server) and on the official EBHI test split if one exists; if no official split exists, use fixed 5-fold cross-validation with the same checkpoint-selection rule as Sec. 4.2. Report the resulting Dice at 224x224 and 512x512 for SAGE and the same-resolution baselines. If the held-out margins over the strongest Table 2 baselines are still positive and the abstract/Table 2 numbers are reconciled into one consistent set, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim is that SAGE-UNet 'achieves superior segmentation on three medical benchmarks' and sets state-of-the-art Dice scores (Abstract; Table 2). The load-bearing condition is that the reported numbers are comparable to prior SOTA numbers. This condition is not met. Section 4.2 states that for EBHI and DigestPath, 'the model checkpoint with the highest validation Dice score was selected, and its validation performance was reported.' Thus the EBHI 95.57% and DigestPath 95.16% numbers are validation scores, not test scores, and they were used for both model selection and final reporting. Published SOTA baselines in Table 2 are generally reported on official test sets; comparing a validation-selected checkpoint's validation score against test-set literature scores biases the comparison in SAGE's favor and there is no error bar or multi-seed estimate to assess the gap. The protocol also changes resolution between tables: Section 4.2 says baseline comparisons are 512x512 while SOTA comparisons are 224x224 (except DigestPath). SAGE-UNet's EBHI Dice is 95.01 at 512 (Table 1) but 95.57 at 224 (Table 2), and DigestPath is 95.28 at 512 (Table 1) but 95.16 at 512 (Table 2), so the reported gains are not stable under the protocol. Additionally, the provided abstract reports 95.23% EBHI / 92.78% and 91.42% GlaS A/B / 91.26% WSI-level DigestPath, while the full-text abstract and Table 2 report 95.57% / 95.16% / 94.17%; these are different numbers with no explanation. Because the central contribution is the empirical SOTA claim, this is a correctness-of-evidence concern, not a style issue. The architectural idea may still be sound, but the paper as written does not establish the claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Shape-Adapting Gated Experts (SAGE), a dynamic-routing framework that augments a static CNN-Transformer U-Net with a dual-path design: a main backbone path and an expert path selected through hierarchical gating (shared vs fine-grained experts, top-K routing), with a Shape-Adapting Hub (SA-Hub) to translate features between CNN and Transformer experts. Evaluations are reported on EBHI, DigestPath, and GlaS. The paper claims state-of-the-art Dice scores of 95.57%, 95.16%, and 94.17%, with additional ablation and routing analyses in the supplement.","tokens_in":16869,"tokens_out":8207,"duration_ms":76115,"significance":"The idea of turning a static hybrid backbone into an input-adaptive expert system is timely and the design is plausible. The paper's strengths include a broad baseline comparison in Table 1, supplementary ablations of the number of selected experts K and shared experts S, and routing visualizations. However, the central SOTA claim is not currently supported: the headline numbers for EBHI and DigestPath are validation-set scores, the SOTA comparison table uses different input resolutions, and the abstract numbers disagree with the full text. If re-evaluated under a sound protocol, the method may still be competitive, but the evidence as presented is insufficient.","major_comments":[{"comment":"For EBHI and DigestPath, the checkpoint with the highest validation Dice is selected and its validation performance is reported. These are validation scores, not test-set scores, and the same validation set is used for both selection and final reporting. This inflates the reported numbers and makes the comparison with published SOTA (which is normally test-set performance) invalid. The abstract's claim that SAGE achieves 95.57% and 95.16% state-of-the-art Dice is not supported without a held-out test evaluation and, ideally, multiple seeds with standard deviations.","section":"§4.2 Evaluation and Model Selection; Table 2"},{"comment":"Baseline comparisons use 512×512 inputs, while SOTA comparisons use 224×224 inputs except for DigestPath. The same SAGE-UNet + Sigmoid model appears in both tables with different results: EBHI DSC 95.01 vs 95.57, DigestPath 95.28 vs 95.16, and GlaS A+B 93.85 vs 94.17. The claimed SOTA margin therefore depends on the resolution and table. A direct comparison to literature numbers is only valid if all methods are evaluated at the same resolution and on the same split. The authors should use one consistent protocol for all comparisons.","section":"§4.2 Comparison Protocol; Tables 1 and 2"},{"comment":"The EBHI evaluation uses only the Adenocarcinoma subset (795 images), but Table 2 gives no indication whether the listed SOTA methods were evaluated on the same subset. If the literature numbers come from the full EBHI dataset or a different class split, the 95.57% figure is not comparable. The authors should clarify this and, if necessary, re-run the SOTA baselines on the same 795-image subset.","section":"§4.1 EBHI Dataset; Table 2"},{"comment":"The abstract on the paper's first page reports 95.23% EBHI, 92.78%/91.42% GlaS Test A/B, and 91.26% WSI-level DigestPath, whereas the full-text abstract and Table 2 report 95.57%, 95.16%, and 94.17% (with GlaS A/B of 94.67%/92.67%). These are materially different numbers. The authors must reconcile them and clearly state which numbers are the canonical results.","section":"Abstract vs. Table 2"}],"minor_comments":[{"comment":"The sentence beginning 'The models based on the transformer structure, such as TransUNet ... can understand long State-space models...' is garbled and seems to conflate transformer and state-space model descriptions. Please rewrite for clarity.","section":"Section 2, Related Work"},{"comment":"Algorithm 1, line 12, omits the output adapter S_out in the expert-path computation, which is inconsistent with Eq. (12)–(13). Please correct the algorithm or clarify the simplification.","section":"Algorithm 1 vs. Eq. (12)–(13)"},{"comment":"The ablation text states that sigmoid surpasses softmax with 95.57% vs 95.05% on EBHI, but Table 2 reports the softmax variant at 94.77%. These numbers should be reconciled.","section":"Supplementary Section 8"},{"comment":"The internal architectures of the shared experts, fine-grained experts, SA-Hub adapters, and the gating networks are not specified. Add these details or release code, otherwise the method is not reproducible.","section":"§4.2 SAGE-UNet Configuration"},{"comment":"The paper motivates SAGE by 'extra computation' and 'efficient resource allocation' but reports no FLOPs, parameter counts, or inference-time comparison. If efficiency is claimed, it should be measured.","section":"§4.3 Quantitative Results"},{"comment":"Each SOTA row lacks a source and a statement of the evaluation split/resolution used to obtain the published number. Add per-row citations and a protocol note.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The validation-set reporting is the main blocker: the EBHI and DigestPath SOTA claims are not verifiable as stated. I would ask the authors to re-run on official test sets where available, or to pre-register a held-out split and report test results with multiple seeds. The abstract discrepancy also needs editorial attention. Given the method is plausible and the baseline comparison is extensive, this seems fixable within a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the architectural idea is real and worth engaging, but the headline SOTA claim doesn't survive contact with the evaluation section. The EBHI and DigestPath numbers in Table 2 are validation scores, selected on that same validation split, and compared to prior published test-set results. That's a biased comparison, and the paper offers no error bars or multi-seed estimates to bracket it.\n\nWhat's genuinely new: the dual-path design with hierarchical gating that first separates shared vs fine-grained experts and then does top-K routing, plus the SA-Hub that adapts feature tensors between CNN and Transformer experts. That combination is not in MoLEx, MoE-NuSeg, or DeepSeekMoE as far as I can tell. The ablations are informative: sigmoid gating consistently beats softmax across configurations, and scaling K and S behaves in a sensible direction. The supplementary routing heatmaps show balanced utilization, though they're qualitative.\n\nThe soft spots are about evidence, not architecture. The biggest is the protocol mismatch. Section 4.2 says for EBHI and DigestPath the checkpoint with highest validation Dice was selected and its validation performance reported. Prior SOTA baselines in Table 2 are typically test-set numbers. Comparing validation-selected validation scores against test-set literature inflates SAGE's apparent gain. Second, resolution shifts between tables: baselines at 512x512, SOTA at 224x224 (except DigestPath), and SAGE's own number moves with resolution — EBHI goes from 95.01 at 512 to 95.57 at 224. That's a big red flag. Third, the abstract's numbers (95.23, 92.78, 91.42, 91.26) don't match Table 2's (95.57, 95.16, 94.17) with no explanation. Fourth, no code or data release, so nothing is independently checkable.\n\nTo be fair, Table 1 compares SAGE against re-implemented baselines at the same 512 resolution and shows consistent gains; that part is internally coherent. And the method is not circular — it's an empirical extension with no invented entities or leakage from what I can see.\n\nWho should read it: people working on MoE for medical image segmentation. The architecture is interesting enough that I'd want to see a clean version with held-out test evaluation, standard test-set comparisons, error bars, and code. As written, the central claim is not established.\n\nI'd send it to peer review — the idea deserves referee attention and the flaws are fixable — but I'd expect major revision and a re-run of the evaluation.","headline":"A plausible dynamic-routing architecture undermined by evaluation protocol: the SOTA numbers are validation scores compared against test-set baselines.","tokens_in":17369,"tokens_out":2115,"would_cite":false,"duration_ms":22734,"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":"SAGE turns fixed CNN-Transformer hybrids into per-input expert routers and reports top Dice scores on three histopathology benchmarks.","keywords":["histopathology segmentation","mixture of experts","dynamic expert routing","CNN-Transformer hybrid","U-Net","whole-slide images","gland segmentation","colorectal cancer"],"falsifier":"Re-run SAGE-UNet and the strongest baselines on the held-out official test sets of EBHI and DigestPath, with all methods at the same input resolution and with checkpoint selection based only on validation; if SAGE's Dice advantage over the best static hybrid falls within noise, the state-of-the-art claim is falsified.","tokens_in":16315,"feed_emoji":"🔬","tokens_out":10175,"duration_ms":96469,"temperature":0.7,"pith_summary":"The paper tries to establish that the standard design of hybrid CNN-Transformer segmentation networks—a fixed, input-agnostic computation graph—is an unnecessary limitation, and that replacing it with per-input expert routing improves accuracy and domain-shift stability. It introduces SAGE, a dual-path layer that keeps the original backbone's transformation as a main path while a hierarchical gate selects among shared and fine-grained experts, with a Shape-Adapting Hub translating feature tensors between convolutional and transformer formats. Instantiated as SAGE-UNet (ConvNeXt plus ViT encoder on a U-Net decoder), the method reports Dice scores of 95.57% on EBHI, 95.16% on DigestPath, and 94.17% on GlaS Test A+B, claiming the best results among compared methods. A sympathetic reader would care because dynamic, input-dependent computation is a plausible way to handle the wide variance in cell size, shape, and tissue appearance in gigapixel histopathology.","feed_headline":"Dynamic expert routing sets new bests on three colon datasets","feed_subtitle":"SAGE lets each patch choose between shared and specialized experts for better accuracy and domain-shift stability.","key_machinery":"The load-bearing machinery is the SAGE block: a dual-path layer whose main path runs the pretrained backbone transformation, and whose expert path executes the top-K selected experts; hierarchical gating, which first uses a scalar gate to bias Semantic Affinity Routing logits between shared and fine-grained experts and then picks the top-K; and the Shape-Adapting Hub (SA-Hub), a pair of learnable adapters that transforms feature shapes so a CNN block can invoke a transformer expert and a transformer block can invoke a CNN expert. The dual-path fusion weight alpha is itself learned, giving the model a per-layer dial between stability and adaptation.","core_discovery":"The paper's central claim is that the static computation graph of hybrid CNN-Transformer networks is a bottleneck: every input is processed through the same fixed sequence of operations regardless of its complexity. SAGE replaces this with a block that routes each feature map through the top-K most semantically relevant experts from a pool of shared and fine-grained specialists, after a high-level gate biases toward either general or specialized computation, and then fuses the expert output with the unchanged main-path output via a learnable scalar. The authors argue that this dynamic restructuring—rather than stacking or concatenating CNN and Transformer modules—explains the measured gains,","pith_inferences":["A testable extension: run SAGE-UNet and the strongest static hybrids on the official EBHI and DigestPath test sets at a single input resolution; the reported validation-selected margins may or may not survive this check.","The SA-Hub's learnable shape adaptation is a generic cross-architecture adapter and could be used to couple other model families (e.g., state-space models) with CNNs or transformers without redesigning the backbone.","The observed layer-wise gating pattern—CNN layers favoring shared experts, deep transformer layers favoring specialized ones—suggests a design heuristic: allocate shared capacity to low-level features and specialization to high-level semantics.","Part of SAGE's measured gain may come from the switch to sigmoid gating rather than shape-adaptive routing; applying sigmoid gating to static hybrid baselines would help separate the two contributions."],"forward_implications":["Static hybrid fusion leaves accuracy on the table: a per-input routing between local and global expert computations is the better operating point.","The hierarchical gating lets the model lean on shared experts for simple or domain-invariant inputs and reserve specialized experts for complex ones, concentrating compute where it is needed.","The SA-Hub demonstrates a general way to mix heterogeneous expert types—CNN, ViT, or others—inside one network, which the authors state is compatible with any CNN or Transformer backbone.","Ablation results give two design rules for MoE-style vision networks: enlarging the number of selected experts (top-K) matters more than adding shared experts, and sigmoid gating outperforms softmax gating.","The large margin on the domain-shifted GlaS Test B suggests dynamic routing reduces over-segmentation and boundary errors when test images differ from training."],"fun_headline_variants":["Dynamic expert routing boosts colon segmentation accuracy","Shape-adapting experts beat static CNN-Transformer hybrids","SAGE: Input-adaptive routing for better WSI segmentation","Letting each patch choose experts improves histopathology segmentation"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The reported state-of-the-art numbers for EBHI and DigestPath rest on the assumption that the checkpoint with the highest validation Dice, evaluated on that same validation split, approximates test-set performance; if validation performance is optimistically biased, the reported margins over prior published results may shrink.","fun_headline_variants_meta":{"raw":{"variants":["Dynamic expert routing boosts colon segmentation accuracy","Shape-adapting experts beat static CNN-Transformer hybrids","SAGE: Input-adaptive routing for better WSI segmentation","Letting each patch choose experts improves histopathology segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000201,"raw_usage":{"total_tokens":1222,"prompt_tokens":754,"completion_tokens":468,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":405}},"tokens_in":498,"tokens_out":468,"duration_ms":5464,"temperature":1.0,"reasoning_tokens":405,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T20:43:10.158111+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run SAGE-UNet and the strongest baselines on the held-out official test sets of EBHI and DigestPath, with all methods at the same input resolution and with checkpoint selection based only on validation; if SAGE's Dice advantage over the best static hybrid falls within noise, the state-of-the-art claim is falsified.","supporting_citations":[],"review_version":1}