{"id":"4275e572-6bf4-49b8-b512-0a5540d66132","arxiv_id":"2505.12074","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A two-branch MIL training method with self-confidence losses and attention calibration improves bag- and instance-level whole-slide image classification on CAMELYON16 and TCGA-NSCLC over WENO and MHIM-MIL baselines.","lead":"The authors propose a training framework for whole-slide image classification that runs bag-level and instance-level branches in parallel and uses each branch's confident predictions to sharpen the other. If the reported gains hold, pathologists and MIL practitioners get a modest accuracy boost on both slide and patch level without extra annotations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Self-confidence losses in Eq. 6 and §3.6 may amplify pseudo-label noise (confirmation bias) rather than correct it; the W2S mechanism is assumed, not demonstrated, and the Conclusion concedes no theoretical guarantee.","rationale":"The reader's weakest assumption correctly identifies the most load-bearing point: the method's novelty and its central 'denoising' claim rest on the self-confidence losses improving pseudo-label quality via weak-to-strong generalization, but the paper provides no argument, and admits in the Conclusion, that this cannot degenerate into confirmation bias. I agree with this assessment after reading the full text: Lself and ℓself are simple self-hardening losses, the two branches share the same encoder and data, and no mechanism distinguishes 'confident correct' from 'confident wrong' predictions. The empirical results are consistent across backbones and datasets and the paper does provide code links and additional encoder experiments, which is real supporting evidence; but consistency alone cannot discriminate the denoising interpretation from entropy minimization or self-training artifacts. I also found an additional specific inconsistency between the main-text ablation and Appendix Table 5, which further weakens the quantitative case for Lself. Since the reader's verdict is already CONDITIONAL and the proposed test would settle the mechanism, I recommend no change in verdict: the paper should be accepted only if the controlled test rules out confirmation-bias amplification.","tokens_in":15268,"tokens_out":7254,"duration_ms":77442,"concrete_test":"Run a controlled noise-injection experiment on CAMELYON16 (C16v) with the ABMIL backbone. At the start of training, corrupt the instance pseudo-labels used in ℓpseudo: for positive bags, flip a known fraction f ∈ {0.1, 0.2, 0.3} of the soft attention labels σ(a_j^i) from their current value to 0, and for negative bags flip the same fraction of zero labels to a positive value. Train the full model and a no-Lself variant (Lself and ℓself removed) under identical schedules and report instance AUC on the test set. If the full model's instance AUC degrades more than the no-Lself variant as f increases, or fails to recover the true labels, then the self-confidence losses amplify rather than correct pseudo-label noise. This directly tests whether the central denoising claim holds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central contribution is that the self-confidence losses Lself (Eq. 6) and ℓself (§3.6) denoise pseudo-labels via weak-to-strong generalization. For this to be true, the model's own confident predictions must be more reliable than the attention-derived pseudo-labels they replace. The paper provides no mechanism enforcing this: both losses harden the current prediction I_t(ŷ) and therefore reward any confident output, including confident errors. This is the classic confirmation-bias failure mode cited in ref [24], which the paper mentions but does not address. The W2S analogy is also not established: W2S requires a stronger student trained under a weak teacher, whereas here both branches share the same encoder, receive the same patch features, and the 'teacher' is the model itself at an earlier epoch. The claimed improvements could therefore be entropy-minimization artifacts rather than evidence of pseudo-label correction. The Conclusion explicitly states there are no theoretical guarantees. The ablation numbers supporting Lself are also internally inconsistent (main text reports drops to 0.9151/0.8915 bag AUC/ACC; Appendix Table 5 gives 0.9134/0.8974, and full instance ACC is 0.9191 vs 0.9291 in Table 1), so the quantitative support for the mechanism is unreliable. Without a controlled test, the central denoising claim is unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a dual-branch multiple instance learning (MIL) framework for whole slide image classification. The bag branch uses attention-based aggregation and a hard positive mining mask, while the instance branch is supervised by soft labels derived from the bag branch's attention scores. The two branches are trained in a scheduled mutual-distillation loop, and both branches include self-confidence losses (Eq. 6 and Section 3.6) that the authors justify through weak-to-strong generalization. The method is evaluated on CAMELYON16 and TCGA-NSCLC with several encoders, and the authors report that their framework improves bag- and instance-level AUC/ACC over the WENO and MHIM-MIL boosting baselines. The central claim is that self-confidence distillation denoises pseudo-labels and yields consistent gains under weak supervision.","tokens_in":15648,"tokens_out":6373,"duration_ms":60283,"significance":"If the central claim is correct, this is a practically relevant contribution to weakly supervised pathology analysis, because it improves instance-level predictions without requiring additional annotation. The evaluation uses external public test labels on CAMELYON16, so there is no derivational circularity in the main results, and the inclusion of both bag- and instance-level metrics is a strength. The reported gains, however, are modest in several cells, the experiments are single-run, and the ablation evidence is internally inconsistent, which weakens the support for the claimed mechanism. The paper does not provide code, which further limits the verifiability of the empirical results.","major_comments":[{"comment":"The ablation results for removing the self-confidence losses are inconsistent between the main text and the appendix: Section 4.4 reports a bag AUC drop to 0.9151 and a bag ACC of 0.8915, while Table 5 reports 0.9134 and 0.8974 for the same setting. Moreover, the full-model instance ACC in Table 5 is 0.9191, whereas Table 1 and the main text report 0.9291. These mismatches mean the quantitative support for the central denoising claim is not reliable as presented; the authors should reconcile the tables and re-verify the numbers.","section":"Section 4.4 and Appendix C, Table 5"},{"comment":"Accuracy is evaluated by optimizing the decision threshold on the test set using Youden's J statistic. This procedure makes the reported ACC values optimistically biased and can change the ranking of methods, especially when ACC differences are small. The authors should report ACC with a threshold selected on the validation set or at a pre-specified operating point, and disclose the selected threshold values.","section":"Section 4.2 (Evaluation Metrics)"},{"comment":"The fully supervised baseline is not specified: the table gives no architecture, no training details, and no statement of which instance-level labels are used for training or validation. Since the text uses this row to claim that the proposed method outperforms fully supervised bag-level classification, this baseline must be fully described and its training protocol must match the weakly supervised setting as closely as possible.","section":"Table 1 (Fully supervised row)"},{"comment":"All results are single-run, and no variance, confidence intervals, or significance tests are reported. The claim of consistent outperformance is not supported by the numbers: in Table 1, TransMIL achieves a bag AUC of 0.9332 on C16, above ABMIL+Ours at 0.9318; in Table 4 with ResNet-18, CLAM-SB+Ours (0.8802 bag AUC) beats ABMIL+Ours (0.8215), and with CONCH, ABMIL+Ours (0.9784) is below plain ABMIL (0.9809). The authors should provide repeated-seed results with confidence intervals and soften the consistency claim accordingly.","section":"Section 4.4 and Tables 1, 4"},{"comment":"The central mechanism of the paper is self-confidence distillation, but the paper does not demonstrate that hardening the model's own predictions corrects pseudo-label noise rather than reinforcing confirmation bias. The weak-to-strong analogy does not directly apply because both branches share the same encoder and receive the same patch features, and the 'teacher' is the model itself at an earlier state. The Conclusion explicitly admits the lack of theoretical guarantees. A controlled experiment, such as tracing pseudo-label accuracy over training or comparing against an entropy-minimization baseline on the same architecture, is needed to support the denoising interpretation.","section":"Section 3.6, Eq. (6) and ℓself"},{"comment":"The hyperparameters β, γ, δ, θ, τ, t, and κ are never reported anywhere in the paper or appendix. Without these values the method cannot be reproduced; the authors must provide them along with the search ranges and, ideally, a sensitivity analysis.","section":"Section 3.6 and Appendix A.3"}],"minor_comments":[{"comment":"The hard positive mining threshold τ appears in the text but its value and the range of the attention scores it is applied to are not defined; please clarify the relationship between σ(ŷ) and the attention scores.","section":"Section 3.4, Eq. (2)"},{"comment":"The notation CE(Ŷ_i, ŷ_i) in L_inst is ambiguous because ŷ_i is a scalar between 0 and 1; the authors should state explicitly that this is binary cross-entropy with ŷ_i as a soft target.","section":"Section 3.6, Eq. (5)"},{"comment":"The DSMIL definition contains an error: 'hm_i = max{h1_i, h2_i, ..., hn_i}' should refer to the max of the instance scores, not the feature vectors; the equation and the surrounding text are inconsistent.","section":"Appendix A.1"},{"comment":"The ablation paragraph says Lself and ℓself are removed, but Table 5 also reports ablations with L_inst and L_attn removed; the text should specify which loss terms are kept in each ablation row.","section":"Section 4.4 (Ablation Study)"},{"comment":"The word 'pseudolable' should be 'pseudo-label', and the reference list appears to be truncated at [46] without a conclusion section; please check the formatting of the bibliography.","section":"Related Work"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the proposed direction, mutual distillation for dual-level MIL, is of interest to the computational pathology community. The main weaknesses are the internal inconsistency in the ablation data and the gap between the strength of the claims and the evidence (single-run results, test-set threshold selection, underspecified baselines). These issues are fixable in a revision, so I do not see grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi [colleague],\n\nQuick take: this is a plausible incremental extension of WENO-style MIL boosting, with consistent gains on two public WSI datasets and several backbones, but the evidence as reported is not clean enough to accept as-is. The novelty is real but modest: the exact combination of bag-instance mutual cross-entropy, self-confidence losses, attention calibration, and periodic scheduling is new, and the paper does the right thing by comparing against WENO and MHIM-MIL using their released code. I also give it credit for testing multiple encoders (ResNet-18/50, CONCH, PLIP) and reporting instance-level results on CAMELYON16, not just bag AUC. The gains are consistent enough across setups that I don't think the headline result is fabricated; the central empirical claim is probably true.\n\nThe soft spots are in the evidence, not the idea. Results are single-run, and ACC thresholds are chosen on the test set via Youden's J, which inflates reported accuracy. The fully supervised baseline in Table 1 has no description of how it was trained, so the \"outperforms fully supervised\" claim is hard to evaluate. More troubling, the main-text ablation for removing Lself gives bag AUC 0.9151 / ACC 0.8915 while Appendix Table 5 gives 0.9134 / 0.8974, and the full-model instance ACC differs between Table 1 (0.9291) and Appendix Table 5 (0.9191). Those numbers need to be reconciled; right now the quantitative support for the key component is unreliable.\n\nThe deeper scientific concern is the W2S story. The self-confidence losses harden the model's own predictions, which is exactly the confirmation-bias failure mode the paper cites in [24]. Since both branches share the same encoder and features, this is not really weak-to-strong teacher-student distillation; it is entropy minimization on the model itself. The paper admits there is no theoretical guarantee. I do not think this sinks the paper, but the mechanism is asserted, not demonstrated.\n\nWho this is for: researchers working on MIL and WSI classification who want a competitive boosting recipe. It is not a conceptual breakthrough, but it is a useful data point. I would send it to peer review and ask for major revision: multiple seeds, validation-based thresholds, code and full hyperparameters, corrected ablation reporting, and ideally a small experiment that tests whether confident errors grow or shrink.","headline":"Plausible incremental MIL boosting recipe with consistent but under-evidenced gains; the weak-to-strong denoising story is asserted rather than demonstrated.","tokens_in":16115,"tokens_out":3014,"would_cite":true,"duration_ms":30722,"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":"A mutual-distillation training loop corrects noisy pseudo-labels in whole-slide image classification and outperforms prior boosting frameworks at both slide and patch levels.","keywords":["multiple instance learning","whole slide image classification","weak-to-strong generalization","knowledge distillation","pseudo-label denoising","digital pathology","self-confidence loss","attention-based MIL"],"falsifier":"Run the full method again but replace every self-confidence target $I_t(\\cdot)$ with the branch's own prediction recorded from a fixed early-training checkpoint, so the pseudo-labels are frozen rather than self-correcting. If instance-level AUC does not drop, the self-correction mechanism is not responsible for the reported gain; if confident predictions become less accurate over training, the loss amplifies confirmation bias.","tokens_in":15088,"feed_emoji":"🔬","tokens_out":9883,"duration_ms":84676,"temperature":0.7,"pith_summary":"Whole-slide image classification in pathology usually trains on slide-level labels only, leaving individual patches unlabeled; methods that add instance-level classifiers often supervise them with noisy pseudo-labels. This paper proposes a training algorithm that lets a bag-level branch and an instance-level branch supervise each other in both directions, and adds self-confidence losses that push each branch toward its own confident predictions, using the weak-to-strong generalization phenomenon to correct pseudo-label errors. The authors claim that this denoising mutual distillation improves both bag-level and patch-level predictions across attention-based MIL backbones, outperforming prior boosting frameworks WENO and MHIM-MIL on CAMELYON16 and TCGA-NSCLC. If correct, the method would give pathologists more accurate tumor localization from slide labels alone, without fine-grained annotations.","feed_headline":"Mutual distillation boosts cancer patch detection from slide labels","feed_subtitle":"Adding self-confidence losses to both branches outperforms prior MIL boosting on pathology benchmarks.","key_machinery":"The central object is a bi-directional distillation loop between two branches sharing one encoder. In the bag branch, attention scores $\\alpha$ are passed through a hard positive mining module $\\mathcal{M}$ that zeroes the normalized attention of any patch whose instance prediction $\\sigma(\\hat{y})$ exceeds a threshold $\\tau$, forcing the bag aggregator to learn from hard patches; the remaining predictions are max-pooled and fed back as bag supervision. A self-confidence loss $L_{\\text{self}} = c_1\\,\\mathrm{CE}(\\hat{Y}, I_t(\\hat{Y})) + c_2\\,\\mathrm{CE}(\\hat{y}^j, I_t(\\hat{y}^j))$ and its instance-branch counterpart $\\ell_{\\text{self}} = \\mathrm{CE}(\\hat{y}^j, I_t(\\hat{y}^j))$ push each branch toward its own confident hard labels, where $I_t$ is an indicator at threshold $t$. The bag loss also includes an attention-calibration term $L_{\\text{attn}}$ that aligns the maximum attention score with the bag label, and the two branches are trained on a periodic schedule ($\\kappa$ bag iterations per instance iteration) to keep them at the same pace. These components together form the tripartite balance the paper argues corrects pseudo-label errors.","core_discovery":"The paper's central claim is that the full model, with bag loss $L_{\\text{bag}} = L_{\\text{label}} + \\beta L_{\\text{inst}} + \\gamma L_{\\text{self}} + \\delta L_{\\text{attn}}$ and instance loss $\\ell_{\\text{inst}} = \\ell_{\\text{pseudo}} + \\theta \\ell_{\\text{self}}$, outperforms WENO and MHIM-MIL boosts at both bag and instance levels. On CAMELYON16, ABMIL+Ours reaches 0.9318 bag AUC and 0.9216 instance AUC; on TCGA-NSCLC it reaches 0.9626 bag AUC. The authors attribute these gains to the self-confidence losses $L_{\\text{self}}$ and $\\ell_{\\text{self}}$, which leverage weak-to-strong generalization to correct pseudo-label noise when the soft labels are not confident or when the bag label and instance supervision disagree. They further state that the method consistently outperforms previous distillation-based boosting methods, especially for instance predictions.","pith_inferences":["The mutual-distillation recipe is likely transferable to other weakly supervised MIL problems beyond pathology, such as video anomaly detection or remote sensing, where pretrained encoders supply the weak-to-strong gap, but the paper does not test this.","Part of the reported gain may come from the larger number of total loss terms rather than from the self-distillation mechanism itself; a controlled comparison that matches loss count would separate these factors.","The threshold $t$ in $I_t$ and the schedule ratio $\\kappa$ are free hyperparameters with unknown sensitivity; tuning them per dataset is a natural extension the paper leaves open.","Because weak-to-strong generalization depends on pretrained representations, the method's advantage may shrink or vanish if the encoder is trained from scratch, which is a testable implication of the stated mechanism."],"forward_implications":["ABMIL+Ours raises bag AUC on CAMELYON16 to 0.9318 and instance AUC to 0.9216, closing most of the gap to fully supervised instance prediction (0.9759) without using patch labels.","On TCGA-NSCLC, the best bag AUC reaches 0.9626, surpassing both WENO and MHIM-MIL boosts on the same backbone.","The improvements are consistent across four attention-based MIL backbones (ABMIL, DSMIL, CLAM-SB, CLAM-MB), indicating the recipe is model-agnostic.","With pathology-specific encoders (CONCH, PLIP), the method still leads on instance-level metrics even when bag-level AUC is saturated above 0.96, showing the main remaining benefit is at the patch level.","Removing the self-confidence losses $L_{\\text{self}}$ and $\\ell_{\\text{self}}$ drops bag AUC from 0.9318 to 0.9151 and instance AUC from 0.9216 to 0.9117 on CAMELYON16, so the paper's reported gains depend on those terms."],"supporting_citations":[{"why":"Defines attention-based MIL aggregation that serves as the backbone for the bag branch and the source of attention-score soft labels.","marker":"[6]"},{"why":"CLAM's instance-level cross-entropy on the top-attended patch inspires the $L_{\\text{label}}$ term in the bag branch.","marker":"[7]"},{"why":"DSMIL provides a dual-stream baseline with an instance classifier and is one of the backbones tested.","marker":"[8]"},{"why":"WENO introduces hard positive instance mining; the paper's hard positive mining module and the comparison against WENO build on this.","marker":"[12]"},{"why":"MHIM-MIL is the masked hard instance mining boosting framework that serves as the main competing baseline.","marker":"[14]"},{"why":"Documents confirmation bias in pseudo-labeling, motivating the need for denoising that the self-confidence losses target.","marker":"[24]"},{"why":"Self-training with noisy student supplies the weak-to-strong generalization idea that the self-confidence losses are designed to exploit.","marker":"[25]"},{"why":"Weak-to-strong generalization is the theoretical-phenomenon basis for treating confident self-predictions as corrective targets.","marker":"[35]"},{"why":"CAMELYON16 provides the bag labels and instance-level ground truth used to evaluate both bag and instance predictions.","marker":"[42]"}],"fun_headline_variants":["Denoising distillation sharpens whole-slide cancer detection","Self-confidence losses fix noisy labels in slide MIL","Weak-to-strong distillation lifts both bag and patch accuracy","Mutual distillation denoises MIL for patch-level diagnosis","Slide-level labels, patch-level accuracy via corrected distillation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that pushing each branch toward its own most confident guesses corrects mistakes in the patch-level labels through the model's ability to generalize beyond its teacher; if that instead locks in the model's early mistakes, the reported gains would be an artifact of training dynamics, and the paper provides no theoretical safeguard.","fun_headline_variants_meta":{"raw":{"variants":["Denoising distillation sharpens whole-slide cancer detection","Self-confidence losses fix noisy labels in slide MIL","Weak-to-strong distillation lifts both bag and patch accuracy","Mutual distillation denoises MIL for patch-level diagnosis","Slide-level labels, patch-level accuracy via corrected distillation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000198,"raw_usage":{"total_tokens":1343,"prompt_tokens":894,"completion_tokens":449,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":372}},"tokens_in":510,"tokens_out":449,"duration_ms":4375,"temperature":1.0,"reasoning_tokens":372,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:41:58.680151+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the full method again but replace every self-confidence target $I_t(\\cdot)$ with the branch's own prediction recorded from a fixed early-training checkpoint, so the pseudo-labels are frozen rather than self-correcting. If instance-level AUC does not drop, the self-correction mechanism is not responsible for the reported gain; if confident predictions become less accurate over training, the loss amplifies confirmation bias.","supporting_citations":[{"cited_title":"Attention-based deep multiple instance learning,","cited_arxiv_id":null,"evidence_quote":"Defines attention-based MIL aggregation that serves as the backbone for the bag branch and the source of attention-score soft labels."},{"cited_title":"Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning,","cited_arxiv_id":null,"evidence_quote":"DSMIL provides a dual-stream baseline with an instance classifier and is one of the backbones tested."},{"cited_title":"Bi-directional weakly supervised knowledge distillation for whole slide image classification,","cited_arxiv_id":null,"evidence_quote":"WENO introduces hard positive instance mining; the paper's hard positive mining module and the comparison against WENO build on this."},{"cited_title":"Multiple instance learning framework with masked hard instance mining for whole slide image classification,","cited_arxiv_id":null,"evidence_quote":"MHIM-MIL is the masked hard instance mining boosting framework that serves as the main competing baseline."},{"cited_title":"Self-training with noisy student improves imagenet classification,","cited_arxiv_id":null,"evidence_quote":"Self-training with noisy student supplies the weak-to-strong generalization idea that the self-confidence losses are designed to exploit."},{"cited_title":"Weak-to-strong generalization: eliciting strong capabilities with weak supervision,","cited_arxiv_id":null,"evidence_quote":"Weak-to-strong generalization is the theoretical-phenomenon basis for treating confident self-predictions as corrective targets."},{"cited_title":"Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer,","cited_arxiv_id":null,"evidence_quote":"CAMELYON16 provides the bag labels and instance-level ground truth used to evaluate both bag and instance predictions."}],"review_version":1}