{"id":"ce3d8fb1-f9f2-4c30-b064-1ee987b97ad5","arxiv_id":"2505.06710","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A weakly supervised pre-training scheme that propagates bag labels to patches improves downstream MIL classification and survival prediction on WSI datasets, but the comparison baselines are not trained on the same target data.","lead":"SimMIL pre-trains the patch feature extractor for whole-slide pathology by copying each slide's cancer label onto all of its patches, then adding augmentation, a prediction head, and a noise-robust loss. Because this weak supervision comes from the same datasets used for evaluation, the reported gains over ImageNet and self-supervised baselines may partly reflect target-domain training rather than a new principle.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SimMIL's central superiority-over-SSL claim is confounded: SSL baselines are pre-trained on NCTCRC while SimMIL uses target-domain weak labels; a domain-matched SSL baseline is missing.","rationale":"The reader's weakest assumption identifies exactly this confound, and I agree. The paper's central claim is comparative; the missing same-domain SSL control is a direct threat to that claim. The fine-tuning and scaling experiments are valuable but do not repair the comparison. The paper is otherwise well-executed, so a conditional verdict remains appropriate. Running the proposed MoCo v2-on-Camelyon16 test would settle whether the observed advantage comes from the weakly supervised scheme or from domain and label access. Until that test is run, the superiority claim should be treated as conditional.","tokens_in":22664,"tokens_out":9245,"duration_ms":86981,"concrete_test":"Pre-train MoCo v2 on the unlabeled Camelyon16 training patches (the same ~2.8M patches and the same MoCo v2 augmentation) for 200 epochs with ResNet18, then evaluate under CLAM-SB exactly as in Table 2. If the AUC gap between SimMIL and MoCo v2 (currently 84.35 vs 78.14) shrinks to overlap within one standard deviation, the domain-match confound is the main driver and the superiority claim is not supported; if the gap persists at the same magnitude, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"SimMIL's headline comparison against SSL is not controlled. In 'Prior arts' (Section 4), MoCo v2 and SimCLR are the released ConCL checkpoints trained for 800 epochs on NCTCRC, a colorectal patch dataset; SimMIL is pre-trained on the training splits of the target datasets (Camelyon16, TCGA-NSCLC, TCGA-BRCA) with weak bag labels for up to 200 epochs. The reported gains (e.g., Camelyon16 CLAM-SB AUC 84.35 vs 78.14 for MoCo v2) therefore conflate the pre-training algorithm with the pre-training data domain and label availability. The preliminary NCTCRC-BAGS study (Table 1) does compare label propagation vs SSL on the same data, and Section 4.3 shows SimMIL fine-tuning improves CTransPath/HIPT, but neither answers whether an in-domain SSL baseline would close the gap. Appendix B uses SSL models pre-trained on 32.6M pathology patches, yet only for linear probing and with a different backbone and pre-training set than the target splits. Without a same-domain, same-backbone SSL baseline, the central claim 'better than self-supervised learning' is under-supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SimMIL, a weakly supervised pre-training framework for the feature extractor in multi-instance learning on whole-slide pathology images. Bag-level labels (or survival risk values) are propagated to all instances, and the extractor is trained with strong augmentation, an MLP prediction head, and a robust loss (SCE for classification, a ranking loss for survival). The method is evaluated on Camelyon16, TCGA-NSCLC, and TCGA-BRCA for classification/subtyping and on TCGA-LUAD, TCGA-BLCA, and TCGA-LUSC for survival prediction, with comparisons to ImageNet pre-training, SSL baselines, and pathology-specific foundation models. Additional experiments study fine-tuning CTransPath/HIPT and scaling to merged datasets. The central claim is that task-specific weak label propagation yields better downstream MIL representations than out-of-domain or task-agnostic pre-training.","tokens_in":22861,"tokens_out":3766,"duration_ms":37534,"significance":"The idea of explicitly encoding the MIL assumption into a pretext task is simple, practical, and likely useful to the computational pathology community. The preliminary NCTCRC-BAGS experiment is well designed and gives an interesting caution about using linear probing and fine-tuning as proxies for MIL representation quality. The compatibility and scaling experiments, if properly controlled, would broaden the contribution. However, the headline comparison against SSL is currently confounded by domain mismatch between the SSL baselines and SimMIL, so the paper's most important empirical claim is not yet convincingly supported.","major_comments":[{"comment":"The comparison against SSL baselines is not controlled. MoCo v2 and SimCLR are the released ConCL checkpoints trained on NCTCRC, a colorectal patch dataset, while SimMIL is pre-trained on the training splits of the target datasets (Camelyon16, TCGA-NSCLC, TCGA-BRCA). The large gains in Table 2, e.g., CLAM-SB AUC 84.35 vs. 78.14 on Camelyon16, could therefore be due to domain match and label availability rather than to the proposed algorithm. I ask the authors to add same-domain, same-backbone SSL baselines pre-trained on the identical target training splits (with comparable epochs), or otherwise explicitly control for the pre-training data domain. This is the load-bearing point for the claim that SimMIL is better than self-supervised learning.","section":"Section 4, 'Prior arts', and Table 2"},{"comment":"Most survival C-index differences are within one standard deviation of the competing methods. For example, with CLAM-SB on TCGA-LUAD, SimMIL reports 59.17±2.51 versus MoCo v2 at 58.23±7.07, and on TCGA-BLCA with ABMIL SimMIL obtains 57.49±7.99 versus MoCo v2 at 58.56±4.04. The text states that SimMIL achieves the best results in almost all experiments, but the reported error bars do not support this level of confidence. The authors should report paired significance tests or per-fold comparisons, or soften the claim.","section":"Section 4.2 and Table 3"},{"comment":"The scaling-law experiment in Table 4 reports a single run without standard deviations, yet several entries are close, e.g., ABMIL on TCGA-NSCLC 89.05 (single) vs. 90.48 (merged), while other entries show very large swings such as DSMIL on TCGA-BRCA Acc 55.15 vs. 89.79. Without repeats or error bars, the claims that merged pre-training consistently outperforms single-dataset pre-training are not established. Please add multiple seeds or otherwise quantify variability.","section":"Section 4.4 and Table 4"},{"comment":"The notation for the SCE loss is confusing: the first term appears to use f(T(x_i)) as a weight multiplying log(Y_i), which is not the standard symmetric cross-entropy formulation used in [85]. Additionally, the ranking loss in Eq. (6) pairs instances from one batch, but it is unclear how comparability and risk ordering are determined when both instances come from the same bag or from different bags with censored outcomes. Please clarify these definitions, since they are central to the method.","section":"Section 3.2, Eq. (5) and Eq. (6)"}],"minor_comments":[{"comment":"The captions contain the typo 'Reuslts' instead of 'Results'.","section":"Table 2 and Table 3 captions"},{"comment":"The description of TCGA-BRCA says 'two subtypes in lung cancer', but the dataset is breast cancer; this is inconsistent with the rest of the sentence.","section":"Section 4, Dataset"},{"comment":"In Eq. (3), the index set is written as j∈{1,...,K}, but K was previously used for the number of bags; the intended range should be the number of instances in bag i, denoted N_i elsewhere.","section":"Section 3.1, Eq. (3)"},{"comment":"The dataset name is written inconsistently as both 'NCTCRC' and 'NCR-CRC-HE-100K'; please standardize the nomenclature.","section":"Section 3.1, preliminary experiments"},{"comment":"For HIPT, the paper states that only the second stage is fine-tuned, but it is not explained why this is sufficient or how the first stage remains compatible; a brief justification would improve clarity.","section":"Section 4.3, Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The empirical core of the paper is promising, but the main SSL comparison suffers from a domain-matching confound that the current manuscript does not resolve. The revision should add in-domain SSL baselines or substantially weaken the 'better than SSL' claim. Also note that the 'first work focusing on representation learning for MIL' statement is too strong given SimpleMIL and prior iterative/pseudo-label MIL works; the authors should soften it. The paper fits the scope of Medical Image Analysis, but the lack of significance testing, especially in the survival tables, will need to be addressed before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things. The paper proposes a simple, practical pre-training recipe for MIL in pathology: propagate bag labels to instance patches, add MoCo-style augmentation, an MLP prediction head, and SCE loss. It works well in their experiments. But the abstract's stronger claim, that this beats self-supervised learning, is not cleanly supported: the SSL baselines (MoCo v2, SimCLR) are released ConCL checkpoints trained on NCTCRC, a colorectal patch dataset, while SimMIL is pre-trained on the target datasets' own training splits. The comparison conflates algorithm, data domain, and label availability.\n\nWhat is genuinely new and good: the assembly, and especially the preliminary NCTCRC-BAGS experiment. Showing that instance-level linear probing and fine-tuning, the usual representation-learning gold standards, do not predict MIL bag-level performance is a real observation worth taking seriously. The compatibility experiments, fine-tuning CTransPath and HIPT with SimMIL, are a practical bonus, and the implementation detail is enough to reproduce the pipeline. The scaling experiment on merged datasets is a nice addition.\n\nThe main soft spot is the SSL comparison. The gains over MoCo/SimCLR, e.g., CLAM-SB AUC 84.35 vs 78.14 on Camelyon16, could shrink or vanish if the SSL models were pre-trained on the same target-domain slides. Table 1 partially mitigates this: SimpleMIL beats SSL on NCTCRC-BAGS using the same pre-training data. But that is SimpleMIL, not the full SimMIL with augmentation, MLP, and SCE, so the gap between the preliminary evidence and the headline claim remains. The comparison against ImageNet is fairer and supports the practical claim that task-specific weak labels help. The \"first work\" statement is an overstatement given iterative MIL fine-tuning methods already in the literature. Minor issues: ablation and scaling tables lack error bars, and in the survival experiments most differences are within one standard deviation.\n\nI would send this to peer review. The core idea is sound, the experiments are extensive, and the preliminary result is a useful methodological signal. The referees should ask for domain-matched SSL baselines on the target training splits, error bars on the ablation tables, and a toned-down novelty claim. I'd bring it to reading group if anyone in our circle works on MIL, and I'd cite the NCTCRC-BAGS observation in work about evaluating representation quality.","headline":"A useful, mostly reproducible pre-training recipe for MIL in pathology, but the headline superiority over self-supervised learning is confounded by domain mismatch in the baselines and needs a controlled follow-up.","tokens_in":23417,"tokens_out":3089,"would_cite":true,"duration_ms":32306,"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":"The paper claims that propagating weak slide-level labels to individual patches during pre-training yields better whole-slide MIL features than ImageNet or self-supervised pre-training.","keywords":["multi-instance learning","weakly supervised pre-training","whole slide image analysis","label propagation","representation learning","computational pathology","survival prediction","self-supervised learning"],"falsifier":"Pre-train MoCo v2 and SimCLR on the same Camelyon16 training slides, with the same patches and budget, that SimMIL uses, then compare downstream CLAM-SB accuracy and AUC on the Camelyon16 test set; if the gap vanishes or reverses, the advantage is access to the target-domain training data, not the weakly supervised pretext task.","tokens_in":22440,"feed_emoji":"🔬","tokens_out":7213,"duration_ms":65179,"temperature":0.7,"pith_summary":"The paper argues that the feature extractor used for whole-slide-image multi-instance learning should be pre-trained with the task's own weak supervision rather than borrowed from ImageNet or learned by task-agnostic self-supervision. Its recipe, SimMIL, propagates each slide's bag-level label to every patch inside the slide, then trains a standard classifier on those noisy per-patch labels while adding strong augmentation, a non-linear prediction head, and a noise-robust loss. Experiments on benign-malignant classification, cancer subtyping, and survival prediction report that this scheme outperforms ImageNet and self-supervised pre-training across attention-based MIL aggregators on breast and lung cancer datasets and three survival cohorts. A sympathetic reader would care because it offers a way to pre-train pathology models using only the slide-level labels that hospitals already have, with no pixel-level annotations.","feed_headline":"Weak slide labels beat ImageNet and self-supervised pre-training","feed_subtitle":"SimMIL pushes bag labels down to patches; downstream MIL classifiers improve across six cancer cohorts.","key_machinery":"The central mechanism is the weakly supervised pretext task built from the standard MIL assumption: the bag label is propagated to every instance, $y_{ij}\\leftarrow Y_i$, and the feature extractor is trained to classify each augmented patch with a symmetric cross-entropy loss $\\mathcal{L}_{\\text{sce}}=-\\sum_{i=1}^{|C|}\\left[\\alpha f(T(x_i))\\log(Y_i)+\\beta Y_i\\log(f(T(x_i)))\\right]$, with a ranking loss $\\mathcal{L}_{\\text{rank}}=-\\sum_{(x_a,x_b)}\\Phi(f(T(x_a))-f(T(x_b)))$ replacing it for survival prediction. The components that make this noisy-label scheme work are strong augmentation $T$, an MLP prediction head $h$ borrowed from BYOL-style architectures, and the robust SCE loss; the prediction head is hypothesized to absorb the distribution of augmented noisy-label inputs and protect the feature extractor from overfitting. This pretext task carries the argument: it injects task-specific, slide-level supervision at the instance level, aligning the inductive bias of the features with the MIL assumption that a bag is positive iff it contains a positive instance, rather than with generic image semantics.","core_discovery":"SimMIL makes the claim that a feature extractor trained by SimpleMIL-style label propagation—every patch in a positive slide is treated as positive, every patch in a negative slide as negative—can serve as a strong pre-training scheme for bag-level MIL, provided three components are added: strong augmentation, a non-linear MLP prediction head, and a symmetric cross-entropy loss for classification or a ranking loss for survival. The paper presents preliminary results on a synthetic bag dataset showing that standard representation-quality metrics such as linear probing and fine-tuning do not predict MIL performance, argues that downstream MIL performance is the right evaluation, and reports that SimMIL's features improve accuracy and AUC over ImageNet and self-supervised baselines across aggregators and tasks. It also shows that fine-tuning pathology-specific self-supervised models with SimMIL for a few epochs improves them, and that pre-training on a merged multi-dataset six-class task scales with data. The underlying insight is that instance-level MIL, long treated as a weak baseline, is better understood as a feature extractor pre-training method for bag-level MIL.","pith_inferences":["A fair test of the central claim would pre-train the self-supervised baselines on the exact same target-training slides as SimMIL; the reported gaps might shrink, which would reframe the contribution as robustness to domain mismatch rather than label-signal superiority.","The noisy-label viewpoint suggests a natural extension: combine SimMIL with explicit noise-robust instance selection or label smoothing, potentially improving performance on slides with low tumor content.","If this scales, hospital archives with slide-level diagnoses could be used directly to pre-train organ-specific models, removing the need for a separate patch-level pre-training corpus.","The ranking-loss variant opens a path to other continuous bag targets, such as measuring tumor burden or treatment response, where an additive instance-level risk model is plausible."],"forward_implications":["Downstream MIL classifiers inherit a feature space whose inductive bias already matches the 'one positive patch makes the slide positive' rule, so attention aggregators focus on tumor patches more sharply.","Pathology-specific self-supervised models can be improved after only 1 to 5 epochs of SimMIL fine-tuning, suggesting weak-label propagation is a cheap final stage for adapting foundation models to slide-level tasks.","Merging slide-level labeled datasets from different cancer sites into one multi-class pre-training task improves downstream performance over single-dataset pre-training at the same data budget, so one shared feature extractor can serve many tasks.","The finding that linear probing and fine-tuning do not predict MIL bag accuracy implies evaluations of pathology representation learning should include downstream MIL performance, not just instance-level transfer metrics.","Survival prediction, where the bag label is a censored risk score, also improves with SimMIL pre-training, so the label-propagation idea extends beyond classification."],"supporting_citations":[{"why":"Supplies the SimpleMIL baseline that SimMIL builds on: direct propagation of bag labels to all instances.","marker":"[22]"},{"why":"Provides the MoCo v2 augmentation recipe of random crop, color jitter, blur, and flip adopted as SimMIL's strong augmentation.","marker":"[43]"},{"why":"Provides the BYOL-style two-layer MLP prediction head architecture used in SimMIL.","marker":"[83]"},{"why":"Introduces symmetric cross entropy, the noise-robust loss SimMIL uses for noisy propagated patch labels.","marker":"[85]"},{"why":"Supplies the DSMIL baseline and the released TCGA-NSCLC patch set, and documents SimpleMIL's failure on low-positive-ratio slides.","marker":"[9]"},{"why":"Provides the CLAM-SB aggregator and dataset preparation protocol used in downstream evaluation.","marker":"[10]"},{"why":"Supplies the NLL-loss survival prediction protocol used to train downstream aggregators.","marker":"[86]"},{"why":"Provides the ranking loss for censored survival data that SimMIL uses as its survival pre-training objective.","marker":"[87]"},{"why":"Releases the MoCo v2 and SimCLR feature extractors pre-trained on a colorectal patch dataset that serve as the SSL baselines.","marker":"[96]"}],"fun_headline_variants":["SimMIL: bag labels train better MIL features than ImageNet","Train MIL features from weak labels, beat ImageNet pretraining","SimMIL pre-training: weak labels outperform self-supervision","Use slide labels to pre-train, MIL improves over baselines","From bag labels to patch features: SimMIL gains on WSIs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes the self-supervised baselines are evaluated on equal footing: their released features were pre-trained on a different patch dataset, while SimMIL was pre-trained on the target datasets' training splits, so the reported gain could partly come from dataset overlap rather than from the label-propagation signal.","fun_headline_variants_meta":{"raw":{"variants":["SimMIL: bag labels train better MIL features than ImageNet","Train MIL features from weak labels, beat ImageNet pretraining","SimMIL pre-training: weak labels outperform self-supervision","Use slide labels to pre-train, MIL improves over baselines","From bag labels to patch features: SimMIL gains on WSIs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000218,"raw_usage":{"total_tokens":1454,"prompt_tokens":971,"completion_tokens":483,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":392}},"tokens_in":587,"tokens_out":483,"duration_ms":5150,"temperature":1.0,"reasoning_tokens":392,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:35:13.275536+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Pre-train MoCo v2 and SimCLR on the same Camelyon16 training slides, with the same patches and budget, that SimMIL uses, then compare downstream CLAM-SB accuracy and AUC on the Camelyon16 test set; if the gap vanishes or reverses, the advantage is access to the target-domain training data, not the weakly supervised pretext task.","supporting_citations":[{"cited_title":"Cheplygina, L","cited_arxiv_id":null,"evidence_quote":"Supplies the SimpleMIL baseline that SimMIL builds on: direct propagation of bag labels to all instances."},{"cited_title":"Grill, F","cited_arxiv_id":null,"evidence_quote":"Provides the BYOL-style two-layer MLP prediction head architecture used in SimMIL."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces symmetric cross entropy, the noise-robust loss SimMIL uses for noisy propagated patch labels."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the NLL-loss survival prediction protocol used to train downstream aggregators."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Releases the MoCo v2 and SimCLR feature extractors pre-trained on a colorectal patch dataset that serve as the SSL baselines."}],"review_version":1}