{"id":"16e1d7a4-edfd-467e-ad87-74599f1ee34d","arxiv_id":"1908.10555","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CAMEL enriches image-level histopathology labels into approximate pixel labels via MIL-based instance selection, and achieves segmentation performance near fully supervised baselines.","lead":"CAMEL is a weakly supervised framework that uses only image-level labels to generate pixel-level segmentation masks for histopathology images, then trains standard segmentation networks on those masks. On two cancer datasets it reports segmentation accuracy close to fully supervised models, which would save pathologists hours of pixel-level annotation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'image-level labels' in Section 4.1 may be derived from pixel-level ground truth, so the experiments do not test the claim that only image-level supervision suffices.","rationale":"The reader's conditional verdict focuses on pseudo-label quality and reproducibility. I agree those matter, but the single most load-bearing assumption is different: the paper's weak labels are not independent of pixel-level ground truth. Section 4.1 states patches are labeled CA if they contain any cancerous region. On CAMELYON16, identifying which 1,280×1,280 patches in a metastatic WSI are CA requires the pixel-level masks; the WSI-level label alone cannot provide this. Thus the 'image-level labels' are derived from the very pixel annotations the method is supposed to avoid. The central claim, as quoted by the reader ('image-level labels alone suffice'), is not tested by the reported experiments. This is a claim-scope problem rather than a mathematical error. It warrants conditional acceptance: the authors should either restate the claim as 'patch-level labels derived from pixel masks' or add an experiment with true WSI-level bags. The concrete WSI-level experiment described above would settle the question. I therefore keep the reader's CONDITIONAL verdict; my concern strengthens the condition but does not move it.","tokens_in":9323,"tokens_out":11854,"duration_ms":133940,"concrete_test":"Run the label-enrichment stage with one bag per WSI instead of one bag per 1,280×1,280 patch, using only the WSI-level CA/NC label as the bag label. Apply the same Max-Max and Max-Min selection to pick representative 1,280×1,280 patches from each WSI, then perform retrain and segmentation exactly as in Section 4.4. If the resulting IoU or F1-score on the CAMELYON16 test set falls materially below the Table 3 CAMEL entries, the claim that image-level labels alone suffice is not supported by the current setup.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 4.1, the 1,280×1,280 training patches are labeled CA/NC by checking overlap with the pixel-level tumor masks; on CAMELYON16 a WSI-level diagnostic label alone cannot determine which patches in a positive WSI are CA. The weak labels used by cMIL are therefore a coarse aggregation of the same pixel-level ground truth used to train the pixel-level FSB. The experiments show that from pixel-derived patch labels the pipeline can regenerate approximate pixel labels and segment comparably; they do not show that CAMEL works from image-level (WSI-level) labels alone. The abstract's central claim is thus broader than the experimental design supports. This is not an internal inconsistency, but it is a load-bearing gap: the motivating scenario of avoiding pixel-level annotation is not actually tested. The pseudo-label accuracy question raised in the reader's weakest_assumption is secondary: Table 2 documents roughly 95% accuracy on the training set, and the downstream segmentation metrics corroborate that the auto-generated labels are adequate. The unresolved issue is upstream, in the provenance of the image-level labels.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes CAMEL, a weakly supervised learning framework for histopathology image segmentation that claims to use only image-level labels. The method first performs label enrichment: it splits images into latticed instances, trains two multiple-instance-learning classifiers using Max-Max and Max-Min selection criteria, combines the selected instances into a fully supervised instance-level dataset, retrains an instance classifier, and relabels all instances. These instance-level labels are then assigned to pixels to produce approximate segmentation masks, which are used to train DeepLabv2 and U-Net in a fully supervised manner. Experiments on CAMELYON16 and a colorectal adenoma dataset report instance-level classification and pixel-level segmentation performance comparable to fully supervised baselines, and the authors state that the adenoma dataset will be released publicly.","tokens_in":9496,"tokens_out":4027,"duration_ms":41710,"significance":"If the central claim were fully validated, CAMEL would be a valuable contribution because it could substantially reduce annotation cost for histopathology segmentation. The paper has several strengths: the method is clearly described; the label-enrichment idea is intuitive; comparisons include multiple weakly supervised baselines and fully supervised upper bounds; and the authors explicitly discuss the over-labeling limitation in Section 5. However, the experimental protocol does not actually test the stated image-level-label setting, and the reported comparisons lack statistical validation. These issues bear directly on the paper's main claim, so the manuscript needs substantive revision before the contribution can be accepted.","major_comments":[{"comment":"The experiments do not actually test the stated 'image-level labels only' setting. In Section 4.1, the 1,280×1,280 training patches are labeled CA/NC by checking whether each patch contains any region marked in the pixel-level tumor masks; the resulting patch-level labels are a coarse aggregation of the very pixel-level ground truth used to train the pixel-level FSB. A WSI-level diagnostic label alone cannot determine which patches in a positive WSI are CA, so the claimed scenario of avoiding pixel-level annotation is not exercised. The abstract and contributions should be revised to describe patch-level weak supervision, or an experiment using true WSI-level labels should be added.","section":"§4.1 and Abstract"},{"comment":"All reported numbers are single-run point estimates with no error bars, confidence intervals, or significance tests. The 'comparable' claim rests on gaps as small as 0.9 percentage points of IoU (85.4 vs 86.3 on CAMELYON16 with DeepLabv2) and 0.2 percentage points of F1 (87.3 vs 87.5 on the adenoma dataset), which could easily be within run-to-run variability. Please report means and variances over multiple runs, or otherwise justify that the differences are stable.","section":"Tables 3 and 4"},{"comment":"The exclusion of Test 114 with the footnote 'because of the duplicate labeling [15]' is a data-selection choice that affects the test set composition. The authors should show that the conclusions are unchanged when Test 114 is included, or provide a principled, pre-specified criterion for its exclusion.","section":"Table 3 footnote"}],"minor_comments":[{"comment":"The quality of the enriched instance labels is evaluated on the same training set from which the pseudo-labels were generated; an independent validation set would strengthen the claim that the pseudo-labels are generally accurate.","section":"§3.1.2 and Table 2"},{"comment":"The weights w1 and w2 are set equal without sensitivity analysis; since w1=w2 is a hyperparameter, a brief ablation would clarify its role.","section":"§3.3.2, Eq. (4)"},{"comment":"The caption defines N = M/m and requires M to be divisible by m, but the text does not state how the scale factor is chosen for the adenoma dataset.","section":"Figure 2 caption"},{"comment":"The phrase 'image-level data' for 1,280×1,280 patches conflicts with the histopathology convention where image-level usually means WSI-level; please define the term explicitly at first use.","section":"§4.1"},{"comment":"There is a typographical error in Section 1: 'theses fine-grained labels' should read 'these fine-grained labels'.","section":"§1"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely of interest to the medical imaging community. The main risk is that the experimental protocol does not match the central claim; if the authors can either run a genuine WSI-label experiment or carefully restrict the claims, I would be willing to see a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read CAMEL (arXiv:1908.10555). The headline claim — that pixel-level segmentation can be trained from image-level labels alone — is not what the experiments test. In Section 4.1, the 1,280×1,280 training patches are labeled CA/NC by checking overlap with the CAMELYON16 pixel-level tumor masks. So the \"weak\" supervision is a coarse aggregation of the same pixel-level ground truth used to train the fully supervised baseline. The method never sees true WSI-level labels. That is a load-bearing gap: the motivating scenario of avoiding pixel-level annotation remains untested.\n\nThat said, the method itself is a reasonable engineering contribution. The cMIL combination of Max-Max and Max-Min selection criteria is a sensible way to counter the selection bias of either criterion alone, and the cascade data enhancement is a clever way to recover information lost when only one instance per bag is used. The reported segmentation results are close to the fully supervised baselines on the datasets used (DeepLabv2 IoU 85.4 vs 86.3 on CAMELYON16; 87.3 vs 87.5 on the adenoma data). The writing is clear, and the over-labeling limitation is acknowledged.\n\nThe other soft spots are real but secondary: results are single-run point estimates with no error bars, Test 114 is excluded post hoc, no code is provided, and the improvement over CDWS-MIL, while positive, is not dramatic. None of these are fatal, but combined with the label-provenance issue, they mean the paper's contribution is better described as \"segmentation from patch-level labels that are derived from pixel masks\" rather than \"segmentation from image-level labels.\"\n\nWho is this for? Someone working on weakly supervised histopathology segmentation would find the cMIL selection and cascade enhancement worth reading, but they should not take the abstract at face value. The paper deserves peer review because the method is nontrivial and the gap is fixable in principle — but a serious referee should require an experiment using true WSI-level labels, and the authors should be asked to reframe the claims.\n\nMy bottom line: worth engaging with, but the central claim needs a major correction.","headline":"Useful MIL-based label enrichment for histopathology, but the 'image-level' supervision in the experiments is actually derived from pixel-level masks, so the central claim is untested.","tokens_in":10101,"tokens_out":2988,"would_cite":false,"duration_ms":28670,"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":"Image-level labels alone can train tumor segmentation models that match pixel-supervised baselines","keywords":["weakly supervised learning","histopathology","image segmentation","multiple instance learning","label enrichment","whole-slide images","cancer detection","CAMELYON16"],"falsifier":"Reproduce CAMEL's enrichment on a held-out cohort whose whole-slide images have independent pathologist pixel annotations, train DeepLabv2 only on the automatically generated masks, and compare against the same network trained on manual masks; the central claim fails if the weakly supervised IoU falls more than roughly two points below the fully supervised IoU, because that would show the pseudo-labels are not close enough to pixel truth to explain the reported parity.","tokens_in":9080,"feed_emoji":"🔬","tokens_out":8559,"duration_ms":78245,"temperature":0.7,"pith_summary":"The paper argues that pixel-level labels, which are expensive in histopathology, may not be necessary: using only image-level labels of whether a tissue patch contains cancer, CAMEL automatically turns each image into a lattice of smaller instances, assigns each instance a label, and then transfers those labels to pixels so that any off-the-shelf segmentation network can be trained. The key claim is that this automatic label enrichment is good enough that the resulting segmentation is comparable to fully supervised training on CAMELYON16 and on a colorectal adenoma dataset, with instance-level classification within about one to two percentage points of a fully supervised baseline. A sympathetic reader would take this as evidence that weak image-level supervision can be converted into strong pixel supervision without pathologist-drawn boundaries, which is what would make large-scale training for cancer segmentation practical.","feed_headline":"Image-level labels alone train tumor segmentation","feed_subtitle":"CAMEL converts whole-slide labels to pixel masks, matching fully supervised models on two datasets","key_machinery":"The load-bearing mechanism is cMIL (combined multiple instance learning), defined by two instance-selection criteria applied to latticed image patches. In multiple instance learning, an image is a 'bag' of patches and only the bag label is known; Max-Max picks the patch with the highest cancer response from every image, while Max-Min picks the highest-response patch from cancer images and the lowest-response patch from non-cancer images. Training two classifiers with these complementary criteria and pooling their selected instances rebalances the pseudo-training data; the method then retrains a fully supervised classifier on those instances, relabels every patch in the original images, and assigns each patch label to its pixels. Cascade data enhancement (two-stage cMIL at different scales) and an image-level constraint loss during retraining are supporting mechanisms that recover information lost by selecting only one instance per image.","core_discovery":"On its own terms, the paper's discovery is that a simple self-training loop can bootstrap pixel-level masks from image-level labels. First, a combined multiple instance learning step splits each image into an $N \\times N$ grid of instances, trains two classifiers that each select one representative instance per image under different criteria (Max-Max and Max-Min), and uses the selected instances, together with the image label, as a pseudo-labeled instance dataset. The classifier is retrained on these pseudo-labels, then applied to every instance of every training image; the resulting instance labels are copied onto all pixels in each instance to form approximate segmentation masks. With these masks, training DeepLabv2 or U-Net produces pixel-level accuracy, F1, and IoU close to models trained on the true pixel annotations. On CAMELYON16, CAMEL (160) reaches an IoU of 85.4 against 86.3 for the fully supervised DeepLabv2 baseline, and the same pattern holds on the colorectal adenoma dataset.","pith_inferences":["The selection-criteria trick is a transferable recipe: use a high-precision positive selector and a high-precision negative selector to build pseudo-labels for any class-imbalanced weakly supervised task, not just histopathology.","Since the paper acknowledges over-labeling (an instance is positive if it contains any cancer pixel), its own logic suggests that adding boundary refinement, such as CRF post-processing or affinity learning, would push CAMEL's segmentation IoU toward or past the fully supervised baseline.","A testable extension is to push the cascade to finer lattices with multiple stages; the paper shows gains from 320 to 160, and the trend suggests that finer granularity may keep improving until the instance size approaches the smallest tumor islands.","The framework's dependence on the retrained classifier's own predictions implies that a poorly calibrated first-stage classifier would propagate errors; measuring enriched-label accuracy on an independent pixel-annotated cohort would tell whether the one-to-two-point gap is stable."],"forward_implications":["A single image-level label can be expanded into $N^2$ instance-level labels, so the cost of annotation drops from pixel boundaries to a yes/no decision per image.","The complementary Max-Max and Max-Min selection is the component that balances sensitivity against specificity; using either criterion alone gives skewed predictions, while the combination approaches the fully supervised trade-off.","Because the enriched labels already support training of standard segmentation architectures, future improvements in segmentation networks can be plugged into CAMEL without changing the labeling step.","Finer instance granularity (160 versus 320) improves segmentation and reduces over-labeling, so further refinement of the lattice is a direct path to closing the remaining gap.","The same label-enrichment pipeline transfers to a second, independently collected histopathology dataset, which supports the paper's claim that the automatic labeling method is general rather than dataset-specific."],"supporting_citations":[{"why":"Supplies the CAMELYON16 dataset, its whole-slide images, and the pixel-level ground truth that define the main benchmark and the fully supervised baselines.","marker":"[1, 5]"},{"why":"Provides the CDWS-MIL baseline and the prior weakly supervised histopathology segmentation approach whose performance CAMEL must beat, as well as the 1280x1280 patch setup that CAMEL adopts.","marker":"[13]"},{"why":"Defines the ResNet-50 architecture used both as the cMIL and retrain classifier and as the shared backbone for the fully supervised instance-level baseline.","marker":"[11]"},{"why":"Supplies the DeepLabv2 segmentation architecture used to train on the enriched pseudo-masks and to produce the reported pixel-level segmentation results.","marker":"[6, 7]"},{"why":"Supplies the U-Net architecture, the second segmentation model used to show that CAMEL's enriched labels work across different fully supervised segmentation networks.","marker":"[20]"},{"why":"Provides the WILDCAT weakly supervised localization and segmentation method used as a natural-image-derived comparison baseline in the segmentation experiments.","marker":"[9]"}],"fun_headline_variants":["CAMEL turns image labels into pixel masks for tumors","Weak labels only? CAMEL still matches full supervision","Pixel-perfect tumor segmentation from slide-level labels","CAMEL: bootstrapping pixel masks from image-level labels","MIL trick: CAMEL rivals fully supervised segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the instance labels automatically produced by a classifier trained on its own pseudo-labels are accurate enough to stand in for pixel-level ground truth, even though any instance containing a speck of cancer is labeled wholly positive (the paper's own over-labeling caveat).","fun_headline_variants_meta":{"raw":{"variants":["CAMEL turns image labels into pixel masks for tumors","Weak labels only? CAMEL still matches full supervision","Pixel-perfect tumor segmentation from slide-level labels","CAMEL: bootstrapping pixel masks from image-level labels","MIL trick: CAMEL rivals fully supervised segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000693,"raw_usage":{"total_tokens":3110,"prompt_tokens":892,"completion_tokens":2218,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":2155}},"tokens_in":508,"tokens_out":2218,"duration_ms":15347,"temperature":1.0,"reasoning_tokens":2155,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:39:45.371278+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce CAMEL's enrichment on a held-out cohort whose whole-slide images have independent pathologist pixel annotations, train DeepLabv2 only on the automatically generated masks, and compare against the same network trained on manual masks; the central claim fails if the weakly supervised IoU falls more than roughly two points below the fully supervised IoU, because that would show the pseudo-labels are not close enough to pixel truth to explain the reported parity.","supporting_citations":[{"cited_title":"Constrained deep weak supervision for histopathology im- age segmentation","cited_arxiv_id":null,"evidence_quote":"Provides the CDWS-MIL baseline and the prior weakly supervised histopathology segmentation approach whose performance CAMEL must beat, as well as the 1280x1280 patch setup that CAMEL adopts."},{"cited_title":"Deep residual learning for image recognition","cited_arxiv_id":null,"evidence_quote":"Defines the ResNet-50 architecture used both as the cMIL and retrain classifier and as the shared backbone for the fully supervised instance-level baseline."},{"cited_title":"U- Net: Convolutional networks for biomedical image segmen- tation","cited_arxiv_id":null,"evidence_quote":"Supplies the U-Net architecture, the second segmentation model used to show that CAMEL's enriched labels work across different fully supervised segmentation networks."},{"cited_title":"WILDCAT: Weakly supervised learning of deep convnets for image classification, pointwise localiza- tion and segmentation","cited_arxiv_id":null,"evidence_quote":"Provides the WILDCAT weakly supervised localization and segmentation method used as a natural-image-derived comparison baseline in the segmentation experiments."}],"review_version":1}