{"id":"2de41676-c08b-4eb7-8cc7-b3a234a78a00","arxiv_id":"2506.12208","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A U-Net-style architecture combining inception-style convolutions with a Mamba state-space block achieves state-of-the-art medical image segmentation at about one-fifth the GFLOPs of the previous best method.","lead":"This paper introduces InceptionMamba, a neural network for segmenting cells, glands, and skin lesions in medical images. It reports slightly better accuracy than prior state-of-the-art models while using about five times less computation, which matters for clinical tools that run on limited hardware.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The core SOTA/efficiency claim depends on protocol-matched comparisons and reproducible GFLOPs; neither is currently demonstrated, so the claim is plausible but unverified.","rationale":"I read the paper as making an empirical SOTA/efficiency claim, so the decisive question is whether the numbers in Tables 1-3 are comparable. The architecture itself is internally coherent: the Inception Mamba module, feature calibration module, and decoder are described in enough detail, and the ablation study shows incremental gains that are consistent with the design choices. I did not find an obvious mathematical flaw in the method. The most load-bearing risk is exactly what the reader identified: comparing against published numbers that may come from different splits, preprocessing, input resolutions, augmentation, or ensembling. The GlaS case is the clearest example, because the paper explicitly describes a five-model ensemble for its own results but does not say which baselines were ensembled. The margins on SegPC21 and ISIC2018 are small, so even modest protocol differences could flip the ranking. The efficiency ratio is also protocol-dependent because GFLOPs depend on input resolution and on how SSM operations are counted; the paper reports neither a FLOPs breakdown nor code to verify it. This concern is real but does not by itself disprove the claim; the authors may have followed earlier protocols faithfully. Therefore the appropriate outcome is to keep the reader's CONDITIONAL verdict, conditional on code release and protocol-matched independent reruns. The duplicate row in Table 4 is a minor reporting issue and does not change the central argument.","tokens_in":11219,"tokens_out":6326,"duration_ms":81228,"concrete_test":"Release code/checkpoints, then download the public GA2-Net code and rerun it under exactly the protocol in Section 3.1: 224x224 input, the same SegPC21 nucleus-cropping preprocessing, the same 290/200/277 split, the same augmentation, and no ensembling. Independently count GFLOPs for both models at 224x224 with the same FLOP counter and same counting rules for depthwise convolutions and the SSM. If GA2-Net's Dice under this matched protocol reaches or exceeds 92.56, the SOTA claim and the 4.85x GFLOPs advantage are likely artifacts; if not, the claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that InceptionMamba beats GA2-Net and previous methods on four datasets at about 4.85x lower GFLOPs. For this claim to hold, the table entries for all compared methods must be produced under the same data split, preprocessing (including the SegPC21 nucleus cropping), input resolution, augmentation, and inference procedure. The paper does not establish this. Section 3.1 sets 224x224 input and says it follows [1,8] for SegPC21 and ISIC2018, and UCTransNet for GlaS, but it does not verify that the published numbers for GA2-Net, UDTransNet, LKMUNet, and others use those exact settings. On GlaS, the text describes three runs of 5-fold cross-validation plus a five-model ensemble during inference; Table 2 does not state which baselines also use ensembling, so the 0.5 Dice lead over SA2-Net could come from ensembling rather than architecture. On SegPC21 and ISIC2018, the margins over the previous best are 0.07 and 0.65 Dice, with no error bars, so a small protocol shift can reverse the ranking. The GFLOPs comparison suffers from the same ambiguity: FLOP counts are only meaningful at a fixed input resolution and with a consistent way of counting SSM and depthwise-convolution operations, and the paper gives no FLOPs profile or code. Without released code/checkpoints and protocol-matched reruns, the 4.85x efficiency advantage and the SOTA accuracy cannot be independently checked.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes InceptionMamba, a hybrid CNN-Mamba segmentation architecture for microscopic and dermoscopic medical images. The model combines a ResNet backbone with a proposed Inception Mamba Module (IMM) and Feature Calibration Module (FCM) in a bottleneck block, plus a lightweight decoder without dense skip connections. The authors report state-of-the-art Dice/IoU on SegPC21, GlaS, ISIC2017, and ISIC2018, and claim approximately 4.85x fewer GFLOPs and 1.45x fewer parameters than the previous best method GA2-Net. The paper includes quantitative comparisons, ablations on SegPC21, qualitative visualizations, and HD95 evaluations.","tokens_in":11518,"tokens_out":4116,"duration_ms":49839,"significance":"If the reported results are protocol-matched and reproducible, the paper offers a practical efficiency-accuracy trade-off for medical segmentation, an area where such trade-offs matter clinically. The ablation study in Table 4 is internally consistent and supports the individual contributions of FCM and IMM. The qualitative results and HD95 scores provide complementary evidence for boundary-quality improvements. However, the central SOTA and efficiency claims are not yet independently verifiable because the comparison protocol across baselines is not established, errors bars are missing on two of the four datasets, and no code or checkpoints are provided. The contribution is incremental (a combination of depth-wise inception-style convolutions with Mamba blocks) but the efficiency claim is substantial if confirmed.","major_comments":[{"comment":"The headline SOTA claim depends on comparing against published baseline numbers that may come from different protocols. The paper fixes the input resolution at 224x224 and states that it follows [1,8] and UCTransNet [27] for splits and augmentation, but it does not verify that the quoted results for GA2-Net, UDTransNet, LKMUNet, U-Mamba, and other baselines were obtained with the same nucleus-cropping for SegPC21, the same resizing, the same augmentation, or the same inference procedure. The margins over the previous best are 0.07 Dice on SegPC21 and 0.65 Dice on ISIC2018, both reported without error bars, so a modest protocol shift could change the ranking. Please report multi-run statistics with standard deviations and, ideally, rerun the strongest baselines under the same code base, or provide explicit evidence that all compared numbers are protocol-matched.","section":"Section 3.1, Tables 1-3"},{"comment":"For the GlaS dataset, the paper states that, following [8,27], it performs three runs of 5-fold cross-validation and uses an ensemble of the five models at inference. Table 2 does not state which of the cited baselines also used ensembling, and several baselines report standard deviations of about 0.4-0.7 Dice. Since the reported Dice lead over SA2-Net is 0.50, the advantage could be attributable to the ensemble rather than to the proposed architecture. Please report single-model results for InceptionMamba as well, or clearly identify the inference protocol used for every row in Table 2.","section":"Section 3.1, Table 2"},{"comment":"The efficiency claim of 4.85x fewer GFLOPs than GA2-Net is only meaningful if all FLOPs are counted with the same conventions and at the same input resolution. The paper reports GFLOPs values but does not describe the counting tool, the inclusion or exclusion of the backbone, or how the SSM scan and depth-wise operations are counted. The ablation table shows GFLOPs values that vary in ways that are not fully explained (e.g., 'IMM with only Mamba' increasing GFLOPs beyond the full IMM). Without code/checkpoints and a stated FLOPs-counting methodology, the efficiency comparison cannot be independently verified. Please release code and provide a detailed FLOPs profile for each component.","section":"Section 2.3, Tables 1 and 4"},{"comment":"The architectural choices that define the final model are selected using the SegPC21 test set. The choice of identity over pooling in IMM, the position of IMM in the decoder, and the number of backbone stages are all validated on the same test set where the final SOTA numbers are reported. This selection-on-test-set procedure can inflate the reported performance. Please add a validation-based model selection protocol, a separate unseen test split, or report the variance across multiple seeds and folds so the reader can assess selection effects.","section":"Section 3.4, Tables 4 and 5"}],"minor_comments":[{"comment":"The table formatting has several textual artifacts, such as '113.1391.10' and 'Dice (%)IoU (%)', where missing spacing makes values run together; please fix the table typesetting.","section":"Table 1"},{"comment":"The HD95 scores for the proposed method are reported without a specification of the distance metric implementation (e.g., symmetric Hausdorff distance, number of surface points sampled) and without HD95 values for all baselines in Tables 1 and 3, so the comparison is incomplete.","section":"Section 3.2"},{"comment":"The description of the FCM input resolutions is ambiguous: the text says the first-stage features are downsampled by a factor of four and the second-stage features by a factor of two before the FCM, but Figure 4 shows a single stride-2 depth-wise convolution; please clarify how these factors are applied.","section":"Section 2.4"},{"comment":"The statement 'we follow [1,8]' for SegPC21 and ISIC2018 is vague because [1] is a general review and [8] is SA2-Net; please list the exact split, preprocessing, and augmentation details for each dataset.","section":"Section 3.1"},{"comment":"The identity branch is said to be 'empirically observed' to outperform the max-pooling branch used in Inception blocks, but no direct ablation comparing identity with max-pooling within the proposed IMM is shown; adding such an ablation would strengthen the claim.","section":"Section 2.3, Table 4"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: InceptionMamba is a U-Net-style segmenter that combines an Inception-style multi-branch depthwise conv, a Mamba selective scan, and a feature calibration module that separates low/high-frequency cues. That combination is new, even though each piece is known. On four public benchmarks (SegPC21, GlaS, ISIC2017, ISIC2018) it reports Dice gains of roughly 0.0–1.2 points over GA2-Net and others, with 4.85x lower GFLOPs than GA2-Net and 1.45x fewer parameters. That efficiency story is the real hook.\n\nWhat it does well: the ablation study is systematic and mostly consistent. Table 4 goes from baseline to full model in sensible steps, and Table 5 tests decoder placement. The FCM design is simple and clearly explained. The authors are candid that they drop skip connections and the fourth backbone stage for efficiency, and the qualitative figures support the boundary-preservation claim. The HD95 numbers, if computed correctly, strengthen the boundary story. The paper is clearly written and easy to follow.\n\nWhere it is soft: the central SOTA/efficiency claim depends on protocol-matched comparisons that the paper does not establish. Section 3.1 says they follow [1,8] and UCTransNet for splits and preprocessing, but the published numbers they compare against may use different input resolutions, ensembling, or preprocessing. On GlaS, they use three runs of 5-fold cross-validation with a five-model ensemble during inference; Table 2 does not say which baselines do the same. A 0.5 Dice lead over SA2-Net could easily come from ensembling. On SegPC21 and ISIC2018 the margins over the previous best are 0.07 and 0.65 Dice, with no error bars in Tables 1 and 3, so a small protocol shift flips the ranking. The GFLOPs comparison is also under-specified: no FLOPs counting methodology, no input-resolution profile for each method, and no code or checkpoints, so the 4.85x advantage is not independently checkable. There is also a duplicated row in the ablation: Exp 7 and Exp 8 both say \"Inception Module [24]\" but give different Dice, which looks like a table typo. Minor, but it should be fixed.\n\nThe stress-test note lands. The core claim is plausible but unverified. Protocol-matched reruns or released code would settle it.\n\nWho this is for: researchers working on efficient medical image segmentation backbones. It deserves a serious referee: the architecture is sensible, the evaluation is broad, and the efficiency claim, if substantiated, is practically useful. I would send it to peer review with a request for code release and error bars, not desk-reject.","headline":"InceptionMamba is a plausible, efficient U-Net variant whose headline SOTA/efficiency numbers rest on protocol-matched comparisons and reproducible FLOPs that the paper does not yet supply.","tokens_in":12096,"tokens_out":2281,"would_cite":false,"duration_ms":25116,"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":"InceptionMamba claims state-of-the-art microscopic and skin-lesion segmentation with about 4.85x fewer GFLOPs and 1.45x fewer parameters than the previous best method.","keywords":["medical image segmentation","microscopic image segmentation","Mamba","selective state space model","feature calibration","Inception module","efficient deep learning","skin lesion segmentation"],"falsifier":"Run InceptionMamba and GA2-Net with identical, scripted preprocessing, augmentation, optimizer, epoch count, and inference on SegPC21 and ISIC2018; if GA2-Net's Dice and IoU then match or beat InceptionMamba, or if the FLOPs gap disappears under the same input resolution, the paper's performance and efficiency claims fail. The single-model, no-ensemble GlaS comparison would be the simplest decisive check.","tokens_in":10989,"feed_emoji":"🔬","tokens_out":7373,"duration_ms":69497,"temperature":0.7,"pith_summary":"InceptionMamba is a medical image segmentation network whose central claim is that multi-stage feature enrichment plus a hybrid Inception-Mamba module reaches state-of-the-art accuracy on four benchmarks while using about 4.85 times fewer GFLOPs and 1.45 times fewer parameters than the previous best method, GA2-Net. The authors' premise is that CNNs and transformers struggle with cluttered backgrounds, blurred boundaries, and large shape and size variation in microscopic images, and that a selective state-space model can supply global context at lower cost. The proposed pipeline enhances features from the first three ResNet stages with a feature calibration module that emphasizes boundary detail, fuses them through a bottleneck containing the Inception Mamba module, and decodes with a lightweight stack that avoids dense skip connections. The reported results on SegPC21, GlaS, ISIC2017, and ISIC2018 support the claim that efficiency and accuracy do not have to trade off. If the comparisons are sound, the practical consequence is that high-accuracy medical segmentation could run on much smaller computational budgets.","feed_headline":"Segmentation model beats SOTA with 5x less compute","feed_subtitle":"InceptionMamba tops four medical datasets with roughly 4.85x fewer GFLOPs and 1.45x fewer parameters than GA2-Net.","key_machinery":"The Inception Mamba module (IMM) is the central object: input features are split channel-wise into several branches, with an identity branch, depthwise convolution branches using square and rectangular kernels of different receptive fields, and a Mamba selective state-space branch, whose outputs are concatenated. The Feature Calibration Module (FCM) smooths features by stride-2 depthwise convolution followed by upsampling, then uses multiplication and subtraction of the original and smoothed maps to emphasize blob regions and fine boundary details. The bottleneck block applies FCM to each of three backbone stages, concatenates and intermixes them with a 1x1 convolution, passes the result through IMM, and adds the IMM output back as a residual. The decoder drops dense skip connections and instead cascades convolutions and upsampling with IMM in the middle, fusing only the stem features at the end.","core_discovery":"The paper's central claim is that channel-wise fusion of Inception-style depthwise convolutions and a Mamba selective state-space block, applied to boundary-enhanced multi-stage features, produces more accurate segmentation of cells and skin lesions than prior CNNs, transformers, and Mamba hybrids, at a fraction of the compute. On SegPC21 it reports 92.56 Dice and 92.37 IoU with a ResNet50 backbone, and 92.84 Dice and 92.63 IoU with a PVT-V2-B2 backbone at 4.86 GFLOPs, against 92.49 Dice and 92.32 IoU for GA2-Net. It also reports lower HD95 boundary error than GA2-Net (1.92 vs 3.87 on SegPC21) and top scores on GlaS, ISIC2017, and ISIC2018. The authors attribute the gains to the feature calibration module, which uses downsampling-upsampling smoothing to separate fine boundary detail from blob structure, and to the Inception Mamba module, whose split branches let the model see multiple kernel shapes and a global selective-scan context simultaneously. The efficiency comes from replacing dense skip connections with a simple decoder and from feeding the Mamba block only a slice of the channels.","pith_inferences":["Extension: the FCM's multiply-and-subtract treatment of original versus smoothed features is a generic low- and high-frequency separation, so the same calibration could be dropped into other dense-prediction heads, such as detection or instance segmentation, without the Mamba branch.","Extension: because the Mamba block sees only one channel subset, the paper implicitly suggests that a partial global context is enough for these datasets; a direct test would be varying the Mamba channel fraction and measuring Dice versus FLOPs, which the paper does not report.","Extension: the reported gains rest on protocol transfer from earlier papers, so re-running all baselines under one scripted pipeline would settle whether the roughly fivefold efficiency advantage is robust or partly an artifact of different preprocessing and ensembling choices."],"forward_implications":["If the comparisons are fair, InceptionMamba gives practitioners a segmentation model that is roughly five times cheaper than the strongest prior method, making high-accuracy cell and lesion segmentation feasible on lighter hardware.","The ablation table implies FCM and IMM are complementary: FCM adds boundary detail, IMM adds multi-context features, and removing either one costs about one Dice point.","The decoder's success without dense skip connections suggests that heavy U-Net-style skip wiring is not necessary once the bottleneck has fused multi-stage features.","Using a PVT-V2 backbone over ResNet50 improves accuracy further and lowers GFLOPs, indicating the framework benefits from a stronger encoder rather than needing custom ones.","Mamba-based global modeling can be restricted to a channel subset and still outperform full transformer encoders, pointing to a cheaper design regime for medical imaging."],"supporting_citations":[{"why":"Supplies the Mamba selective state-space block that gives IMM its global-context branch.","marker":"[10]"},{"why":"U-Net is the baseline architecture and the main comparison standard for the decoder and bottleneck design.","marker":"[21]"},{"why":"GA2-Net is the previous best method this paper compares against for the 4.85x GFLOPs and 1.45x parameters claim.","marker":"[9]"},{"why":"SA2-Net provides the SegPC21 and ISIC evaluation protocol (nucleus cropping, splits, augmentation) and is a key comparison baseline.","marker":"[8]"},{"why":"UCTransNet supplies the GlaS training protocol, including batch size, learning rate, and three-times five-fold ensembling, that the paper follows.","marker":"[27]"},{"why":"The Inception block motivates the multi-kernel depthwise convolution design of the IMM.","marker":"[23]"},{"why":"ResNet50 is the ImageNet-pretrained backbone whose stem and first three stages provide the multi-scale features being enriched.","marker":"[12]"},{"why":"GlaS is one of the four evaluation benchmarks and defines the gland segmentation task used in the comparison.","marker":"[22]"}],"fun_headline_variants":["InceptionMamba: SOTA segmentation with 5x less compute","Mamba-Inception hybrid beats prior best on four datasets","Efficient cell segmentation: 5x compute cut, top Dice scores","State-space model for micro-images: SOTA with 5x fewer FLOPs","InceptionMamba sharpens cell boundaries, cuts compute 5x"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim assumes every compared method was trained and tested under the same protocol, meaning the same splits, preprocessing (nucleus cropping and 224x224 resizing), augmentation, loss, and the five-fold ensemble used for GlaS, because the paper follows prior papers' settings rather than retraining all baselines under one pipeline.","fun_headline_variants_meta":{"raw":{"variants":["InceptionMamba: SOTA segmentation with 5x less compute","Mamba-Inception hybrid beats prior best on four datasets","Efficient cell segmentation: 5x compute cut, top Dice scores","State-space model for micro-images: SOTA with 5x fewer FLOPs","InceptionMamba sharpens cell boundaries, cuts compute 5x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000384,"raw_usage":{"total_tokens":2099,"prompt_tokens":1082,"completion_tokens":1017,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":698,"completion_tokens_details":{"reasoning_tokens":921}},"tokens_in":698,"tokens_out":1017,"duration_ms":69823,"temperature":1.0,"reasoning_tokens":921,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:56:17.884262+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run InceptionMamba and GA2-Net with identical, scripted preprocessing, augmentation, optimizer, epoch count, and inference on SegPC21 and ISIC2018; if GA2-Net's Dice and IoU then match or beat InceptionMamba, or if the FLOPs gap disappears under the same input resolution, the paper's performance and efficiency claims fail. The single-model, no-ensemble GlaS comparison would be the simplest decisive check.","supporting_citations":[{"cited_title":"U-net: Convolutional networks for biomedical image segmentation","cited_arxiv_id":null,"evidence_quote":"U-Net is the baseline architecture and the main comparison standard for the decoder and bottleneck design."},{"cited_title":"Guided-attention and gated-aggregation network for medical image segmentation","cited_arxiv_id":null,"evidence_quote":"GA2-Net is the previous best method this paper compares against for the 4.85x GFLOPs and 1.45x parameters claim."},{"cited_title":"SA2-Net: Scale-aware Attention Network for Microscopic Image Segmentation","cited_arxiv_id":"2309.16661","evidence_quote":"SA2-Net provides the SegPC21 and ISIC evaluation protocol (nucleus cropping, splits, augmentation) and is a key comparison baseline."},{"cited_title":"Uctransnet: rethinking the skip connections in u-net from a channel-wise perspective with transformer","cited_arxiv_id":null,"evidence_quote":"UCTransNet supplies the GlaS training protocol, including batch size, learning rate, and three-times five-fold ensembling, that the paper follows."},{"cited_title":"Gland segmentation in colon histology images: The glas challenge contest","cited_arxiv_id":null,"evidence_quote":"GlaS is one of the four evaluation benchmarks and defines the gland segmentation task used in the comparison."}],"review_version":1}