{"id":"4b3beac2-10c0-40a7-b050-c84bb18fb37d","arxiv_id":"2412.16711","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Pixel-Mamba, an end-to-end Mamba-based architecture with progressive token expansion, reports tumor staging and survival scores on three TCGA datasets that match or exceed several pathology foundation models without pathology-specific pretraining.","lead":"Pixel-Mamba treats every pixel of a whole slide image as a token, then progressively merges neighboring tokens into larger regions while a state space model maintains global context. Pretrained only on ImageNet, it reports higher tumor staging and survival scores than several pathology foundation models on three TCGA cancer datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claim of surpassing SOTA FMs is under-supported: UNI/CHIEF omitted and flagship results tuned on BLCA test performance.","rationale":"Pixel-Mamba is a coherent architecture, and the ablations support the value of hierarchical token expansion; the ImageNet results are plausible and not the problem. The load-bearing step is the comparison. A paper claiming to surpass SOTA FMs cannot omit the strongest FMs, especially when it cites them. UNI and CHIEF are established WSI foundation models and were available before submission. Their absence leaves the 'SOTA' quantifier vacuous. Additionally, the reported margins over the included baselines are often smaller than one standard deviation, so even the included comparison does not demonstrate superiority. The α=0.8 selection on BLCA in Appendix A.3 further makes the headline BLCA result a post-selected number; I agree with the reader's concern about this, but I see the more decisive missing evidence as the absence of UNI/CHIEF and the lack of significance testing, rather than only the faithfulness of the reimplementation. I would therefore keep the verdict conditional, with the explicit condition that the SOTA comparison be expanded and that hyperparameters be selected without using test folds.","tokens_in":17174,"tokens_out":9332,"duration_ms":81970,"concrete_test":"Reproduce Table 2 on the same TCGA 5-fold splits using the official UNI and CHIEF encoders with the same MIL heads (ABMIL/TransMIL/ILRA-MIL), and re-select α on a held-out validation split of BLCA rather than on the BLCA evaluation folds. If either UNI or CHIEF attains the best macro-F1/C-index on any task, or if the validation-selected α is not 0.8 and materially lowers BLCA results, the central superiority claim fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is superiority over SOTA pathology foundation models without pathology-specific pretraining. This requires comparing against the strongest FMs, yet Table 2 omits UNI and CHIEF, both cited in §2.2 as typical FMs and both pretrained on hundreds of thousands to millions of WSIs/patches at submission time. The included GigaPath/CONCH/HIPT baselines do not establish 'SOTA.' Even among included baselines, reported advantages are small relative to cross-validation spread: BLCA survival C-index 0.6507±0.0485 vs CONCH+TransMIL 0.6404±0.0253; BLCA staging macro-F1 0.5334±0.0608 vs GigaPath+ILRA-MIL 0.5320±0.0487. Appendix A.3 selects α=0.8 by maximizing C-index on BLCA, so the flagship BLCA result is an optimized point rather than an independent evaluation. The combination of missing strongest baselines, test-set hyperparameter selection, and wide error bars means the abstract's 'surpasses SOTA' claim is not currently supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Pixel-Mamba, an end-to-end state-space model for gigapixel whole-slide image analysis. Raw pixel tokens are serialized with a region-based zigzag scan; a bidirectional Mamba block provides global context; token expansion progressively enlarges receptive fields from 1x1 to 32x32; and a region-fusion module merges similar regions via CLS-token cosine similarity. The authors report ImageNet-1K classification and TCGA tumor staging and survival results, claiming that Pixel-Mamba, pretrained only on ImageNet, matches or exceeds foundation-model-based two-stage MIL pipelines without pathology-specific pretraining.","tokens_in":17474,"tokens_out":3817,"duration_ms":34018,"significance":"If the central comparison were fully supported, Pixel-Mamba would be a valuable lightweight end-to-end baseline for computational pathology, with linear-memory sequence modeling and a clear hierarchical design. The architecture is described in detail, and the ablations (Tables 3-6 and Table 9) usefully isolate the contributions of hierarchical tokenization, scan window, magnification, and the retention rate alpha. The ImageNet-1K results also show that Pixel-Mamba is competitive with existing SSM backbones at modest parameter counts. However, the main empirical claim of surpassing state-of-the-art pathology foundation models depends on benchmark comparisons that are currently undermined by missing baselines, in-sample hyperparameter selection, and overlapping error bars.","major_comments":[{"comment":"The central claim that Pixel-Mamba 'achieves or surpasses' SOTA pathology foundation models is not supported because the strongest contemporary foundation models, UNI and CHIEF, are absent from Table 2 even though both are cited in §2.2 as typical pathology FMs. GigaPath, CONCH, and HIPT are strong baselines, but they do not by themselves establish that the comparison is against SOTA. The authors should add UNI and CHIEF results or temper the abstract claim to say 'the evaluated foundation models.'","section":"Abstract and §4.3, Table 2"},{"comment":"The flagship BLCA result is partly a fitted result. Appendix A.3 selects alpha = 0.8 by maximizing the C-index on BLCA, and Table 5 selects the scan-window size of 224 by evaluating on the same BLCA benchmark. These hyperparameters are then used for all downstream experiments, so the reported BLCA numbers are not an independent evaluation of the configuration. The authors should either use a held-out validation set or nested cross-validation for hyperparameter selection, or explicitly report the selection procedure and its impact on the main claims.","section":"Appendix A.3, Table 9, and §4.5, Table 5"},{"comment":"The reported advantages over two-stage baselines are often within one standard deviation, making the claim of 'outperforming' statistically unsupported. For example, BLCA survival C-index is 0.6507±0.0485 for Pixel-Mamba vs 0.6404±0.0253 for CONCH+TransMIL; BLCA staging macro-F1 is 0.5334±0.0608 vs 0.5320±0.0487 for GigaPath+ILRA-MIL. The paper provides no significance tests across folds or patient-level bootstrap intervals. The authors should add appropriate statistical tests and discuss whether these differences are meaningful.","section":"Table 2 and §4.3/§4.4"},{"comment":"The two-stage baseline results need more implementation detail to be verifiable and fair. The text does not state the patch size, tissue-detection preprocessing, magnification, or whether official pretrained weights and feature-extraction code were used for GigaPath, CONCH, and HIPT. Since these FMs are known to be sensitive to patch-extraction protocol, the authors should document the exact feature-extraction pipeline and, if possible, release the extracted features or use official embeddings.","section":"§4.1 and Table 2"}],"minor_comments":[{"comment":"The notation in Eq. (1) is inconsistent: f_c is written as 'f c' in the text and the subscripts f_cx and f_cz are not defined. Please clarify which linear layers are shared and which are distinct, and define norm.","section":"Eq. (1), §3.3"},{"comment":"The text refers to 'GigaViT' in the staging-results paragraph, but the table and related work call it GigaPath. Please use consistent names.","section":"§4.3"},{"comment":"The LongViT rows are labeled 0.6x, 2.5x, and 5.0x without explaining what these scaling factors mean in the main text. Please define them or refer to the supplementary material.","section":"Table 2 and §4.1"},{"comment":"The serialization step says each pixel is tokenized as an RGB token, but the input normalization (e.g., whether pixel values are scaled to [0,1]) is not described in the implementation details. Please specify the input preprocessing.","section":"§3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is well positioned for a computer-vision venue but the current comparison does not yet substantiate the 'surpasses SOTA foundation models' claim. The missing UNI/CHIEF baselines and the BLCA-based selection of alpha and scan window are the two issues most likely to be raised by reviewers. If the authors can provide a clean evaluation with held-out hyperparameter selection and stronger baselines, the paper could be a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Good to know: this is a real architecture paper, not a padding job. Pixel-Mamba does something the prior work didn't: it starts from raw pixels, expands tokens hierarchically from 1x1 to 32x32 through a bidirectional Mamba, and uses region fusion to keep memory manageable. The ablation story is coherent — hierarchical tokenization, scan window, and input magnification all move the C-index on BLCA in the expected directions. The ImageNet numbers (77.8 and 80.8) show the backbone is competitive with Vim at similar size, so the engineering is legitimate.\n\nThe problem is the central claim in the abstract: that this matches or beats SOTA foundation models pretrained on millions of WSIs. The comparison as run doesn't support that. First, the two strongest FMs the paper itself cites — UNI and CHIEF — are absent from Table 2. GigaPath and CONCH are good but they are not the current bar. Second, the hyperparameter alpha=0.8 was chosen by maximizing C-index on BLCA (Appendix A.3), and the scan window 224 is justified from pretraining. So the flagship BLCA result is an optimized point, not an honest evaluation. Third, the error bars are wide — BLCA survival 0.6507±0.0485 vs CONCH+TransMIL 0.6404±0.0253; that difference is noise. There's no significance test comparing models, only a Log-Rank test of Pixel-Mamba's own risk groups.\n\nThat said, the architecture is worth taking seriously. The ablations are informative, the region fusion idea is clever, and the paper honestly reports its limitations (2.5x magnification, no pathology pretraining). It's not a fraud; it's an overclaimed engineering contribution. If the authors add UNI/CHIEF baselines, do hyperparameter selection on a validation fold instead of the test benchmark, and add proper significance testing, the claim may survive. As is, I'd advise the editor to send it to review, because the core idea deserves scrutiny, but the authors need to fix the evaluation before publication.","headline":"Genuinely new end-to-end SSM architecture for gigapixel WSIs, but the headline claim of surpassing SOTA foundation models is not supported by the current evaluation.","tokens_in":17978,"tokens_out":2651,"would_cite":true,"duration_ms":22959,"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":"Pixel-Mamba claims gigapixel whole-slide analysis can match pathology foundation models without any pathology-specific pretraining.","keywords":["whole slide images","state space model","Mamba","end-to-end training","hierarchical representation","computational pathology","tumor staging","survival analysis"],"falsifier":"Reproduce the paper's protocol but select alpha, scan window, and magnification on held-out validation for each task separately, and run the two-stage baselines with their original published hyperparameters; if Pixel-Mamba's margins vanish or reverse, the claim that it matches foundation models without pathology pretraining fails.","tokens_in":17022,"feed_emoji":"🔬","tokens_out":3030,"duration_ms":26494,"temperature":0.7,"pith_summary":"Pixel-Mamba is an end-to-end architecture that consumes whole slide images at the pixel level and builds slide-level representations with Mamba, a state-space model whose memory cost grows linearly with sequence length. The paper's central claim is that by progressively expanding token receptive fields from 1x1 to 32x32 and fusing similar regions, the model captures both local and global hierarchy in gigapixel images. On tumor staging and survival analysis across three TCGA cancer datasets, Pixel-Mamba matches or beats two-stage MIL pipelines that rely on large pathology foundation models, even though Pixel-Mamba was pretrained only on ImageNet. If correct, this means pathology-specific pretraining at massive scale is not required for competitive slide-level analysis, and end-to-end training on raw pixels is a practical baseline.","feed_headline":"Gigapixel pathology model trains end-to-end on raw pixels","feed_subtitle":"A 6.2M-parameter state-space model beats two-stage pipelines on tumor staging and survival without pathology pretraining.","key_machinery":"The central mechanism is progressive token expansion combined with region fusion inside a Mamba state-space backbone. Pixel tokens are serialized with a region-based zigzag scan, then each layer runs a bidirectional Mamba block to model long-range dependencies, while region fusion merges the most similar regions using CLS-token cosine similarity to cut redundancy and memory. Token expansion alternately concatenates and averages neighboring tokens along horizontal and vertical axes, growing the receptive field from 1x1 to 32x32 across 24 layers, so the network builds hierarchical multi-scale representations while keeping end-to-end training feasible on gigapixel inputs.","core_discovery":"The paper claims that an end-to-end state-space model can outperform two-stage multiple-instance-learning pipelines that use pathology foundation models, without any pathology-specific pretraining. Concretely, Pixel-Mamba-Stage achieves macro F1 scores of 0.5334, 0.3744, and 0.3917 on BLCA, BRCA, and LUAD tumor staging, while Pixel-Mamba-Surv achieves C-index values of 0.6507, 0.6707, and 0.6468 on survival analysis for the same datasets. These numbers exceed the reported results of two-stage MIL with GigaPath, CONCH, and HIPT features, as well as the end-to-end LongViT baseline, with a 6.2M-parameter backbone.","pith_inferences":["The comparison may be sensitive to tuning asymmetries: the paper selects its own region-retention alpha and scan window from BLCA validation results, so an independent evaluation with per-task validation selection would clarify whether the margin reflects architecture or hyperparameter choice.","If the end-to-end advantage generalizes, it implies that spurious correlations in patch-level features, which two-stage pipelines inherit from pretrained encoders, can be corrected by jointly optimizing the whole slide representation.","A natural extension is to test Pixel-Mamba on molecular marker prediction or cancer subtyping tasks, where foundation-model embeddings currently dominate, to see whether raw-pixel end-to-end learning transfers beyond staging and survival.","The hierarchical token schedule from 1x1 to 32x32 could be transferred to other dense prediction domains, such as satellite or aerial imagery, where similar multi-scale local-global tradeoffs arise."],"forward_implications":["If the claim holds, end-to-end state-space models become a competitive baseline for whole-slide analysis without needing expensive pathology foundation-model pretraining.","The linear memory scaling of Mamba could make end-to-end training on larger magnifications, such as 10x or 20x, tractable with modest GPU resources.","Progressive token expansion offers a general recipe for injecting locality into long-sequence vision models, potentially benefiting other gigapixel or high-resolution imaging tasks.","Region fusion based on CLS-token similarity suggests that redundancy reduction can be learned dynamically rather than fixed by patch grids."],"supporting_citations":[{"why":"Introduces Mamba, the state-space model that gives Pixel-Mamba its linear memory complexity and selective state-space backbone.","marker":"[12]"},{"why":"Vision Mamba provides the bidirectional SSM block and ImageNet pretraining recipe that Pixel-Mamba adapts.","marker":"[54]"},{"why":"LongViT is the main end-to-end WSI baseline that Pixel-Mamba must beat on pathology tasks.","marker":"[41]"},{"why":"GigaPath is the vision foundation model whose patch features feed the two-stage MIL baselines compared against Pixel-Mamba.","marker":"[48]"},{"why":"CONCH is the vision-language foundation model whose patch features feed the two-stage MIL baselines compared against Pixel-Mamba.","marker":"[27]"},{"why":"HIPT is the hierarchical two-stage representation method that Pixel-Mamba compares against on staging and survival.","marker":"[4]"},{"why":"ABMIL is one of the attention-based multiple-instance-learning aggregators used in the two-stage baselines.","marker":"[19]"},{"why":"TransMIL is a transformer-based MIL aggregator used as a two-stage baseline.","marker":"[36]"},{"why":"ILRA-MIL is a low-rank MIL aggregator used as a two-stage baseline and highlighted in the Kaplan-Meier comparison.","marker":"[44]"},{"why":"LocalMamba motivates the scan-window mechanism that Pixel-Mamba uses to keep spatially adjacent tokens close in the sequence.","marker":"[16]"}],"fun_headline_variants":["State-space model beats pathology foundation models without pretraining","End-to-end Mamba analyzes gigapixel WSIs, beats two-stage MIL","Pixel-Mamba: 6.2M params, no pretraining, SOTA on WSI tasks","End-to-end Mamba matches pretrained foundation models on WSIs","State-space model for gigapixel pathology, no pretraining needed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central comparison is fair only if the reimplemented two-stage baselines were configured as carefully as Pixel-Mamba; the authors select their own alpha and scan window from the BLCA validation set, so if the baselines were not tuned similarly, the reported margin could shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["State-space model beats pathology foundation models without pretraining","End-to-end Mamba analyzes gigapixel WSIs, beats two-stage MIL","Pixel-Mamba: 6.2M params, no pretraining, SOTA on WSI tasks","End-to-end Mamba matches pretrained foundation models on WSIs","State-space model for gigapixel pathology, no pretraining needed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000758,"raw_usage":{"total_tokens":3352,"prompt_tokens":912,"completion_tokens":2440,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":2341}},"tokens_in":528,"tokens_out":2440,"duration_ms":15025,"temperature":1.0,"reasoning_tokens":2341,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:20:03.081863+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce the paper's protocol but select alpha, scan window, and magnification on held-out validation for each task separately, and run the two-stage baselines with their original published hyperparameters; if Pixel-Mamba's margins vanish or reverse, the claim that it matches foundation models without pathology pretraining fails.","supporting_citations":[{"cited_title":"A whole-slide foundation model for digital pathology from real-world data","cited_arxiv_id":null,"evidence_quote":"GigaPath is the vision foundation model whose patch features feed the two-stage MIL baselines compared against Pixel-Mamba."},{"cited_title":"A visual- language foundation model for computational pathology","cited_arxiv_id":null,"evidence_quote":"CONCH is the vision-language foundation model whose patch features feed the two-stage MIL baselines compared against Pixel-Mamba."},{"cited_title":"Scaling vision transformers to gigapixel images via hierarchical self-supervised learning","cited_arxiv_id":null,"evidence_quote":"HIPT is the hierarchical two-stage representation method that Pixel-Mamba compares against on staging and survival."},{"cited_title":"Attention-based deep multiple instance learning","cited_arxiv_id":null,"evidence_quote":"ABMIL is one of the attention-based multiple-instance-learning aggregators used in the two-stage baselines."},{"cited_title":"Transmil: Transformer based correlated multiple instance learning for whole slide image classification","cited_arxiv_id":null,"evidence_quote":"TransMIL is a transformer-based MIL aggregator used as a two-stage baseline."},{"cited_title":"Exploring low-rank property in multiple instance learning for whole slide image classifica- tion","cited_arxiv_id":null,"evidence_quote":"ILRA-MIL is a low-rank MIL aggregator used as a two-stage baseline and highlighted in the Kaplan-Meier comparison."}],"review_version":1}