{"id":"99c3ebe0-6d86-427a-8f80-dceaf79aa8d2","arxiv_id":"2411.13528","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Entropy of a Bayesian network trained on sparse point labels is proportional to nucleus probability, enabling weakly supervised nuclei detection that reaches 0.724 mAP50 versus 0.834 for fully supervised training.","lead":"Point labels, one dot per nucleus, are enough to train a model that finds and outlines nuclei almost as well as one trained on full hand-drawn contours. The method uses the entropy of a Bayesian segmentation network as a stand-in for the full mask, then feeds those masks to a standard instance segmentation model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Appendix A derivation assumes a constant per-pixel label probability ε, but the implemented random-point-plus-radius sampling makes ε strongly dependent on nucleus size and boundary distance; Eq. (22)'s proportionality H ∝ P(CT) is therefore not established by the stated procedure.","rationale":"The paper's most distinctive contribution is the theoretical justification that the entropy of the Bayesian network output is proportional to P(CT) (Eq. 22). I read Appendix A carefully. The derivation concerns the entropy of the label distribution H_binary(P(CL)), where P(CL)=ε P(CT) under the assumption that ε is a fixed constant and background is never spuriously labeled. The actual sampling described in Section 3.2 and Appendix B is one random point per nucleus plus a radius-3 disk. That procedure makes the per-pixel label probability ε(x) = area(B(x,3)∩nucleus)/A, which varies by more than an order of magnitude with nucleus size and is also position-dependent inside a nucleus; disks near the boundary also label background, contradicting P(CL|C_cT)=0. Since the exact entropy of the label at a true-nucleus pixel is H_binary(ε(x)), the clean proportionality Eq. (22) does not follow. The reader's flagged assumption that the network is a good entropy estimator is also valid, but even granting it, the target entropy is not −ε log ε P(CT) under the real sampling. The empirical Dice/AUROC show good separability, which is consistent with a merely monotone relationship, but do not validate the specific constant-ε model. I also note Eq. (21) in the proof is algebraically incorrect: the limit is −1/log x, not 0, though the qualitative dominance of −ε log ε survives. The method may still work as a heuristic; the detection results are plausible. But the theoretical centerpiece needs revision or supporting simulation, so the paper should remain CONDITIONAL on addressing this gap.","tokens_in":12087,"tokens_out":14933,"duration_ms":151565,"concrete_test":"Run a Monte Carlo simulation of the label-generation process on the PanNuke ground-truth masks (or a synthetic set of nuclei with areas spanning 30–5000 px): for each nucleus, sample a large number (e.g., 1000) of uniform random points and add the radius-3 disk; record the empirical per-pixel label frequency for all pixels. Report the distribution of ε(x) stratified by nucleus area and by distance to boundary, plus the fraction of background pixels labeled (spillover). If the coefficient of variation of ε(x) within true nuclei is large (e.g., >20%) or if ε(x) correlates with nucleus area, Eq. (22) is not a valid description of the labels; the authors should replace the proportionality claim with a monotonicity argument or change the sampling to achieve a uniform ε.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim is Eq. (22): H ≈ −ε log ε P(CT). It is derived in Appendix A from P(x_L=Nucleus | x_T=Nucleus)=ε (Eq. 13) and P(x_L=Nucleus | x_T=Background)=0 (Eq. 16). These are not properties of the stated labeling procedure. Section 3.2 samples one random pixel per nucleus and adds a radius-3 neighborhood. For a true-nucleus pixel x inside a nucleus of area A, the conditional label probability is ε(x)=area(B(x,3)∩nucleus)/A, not a constant. For a small nucleus (A ≈ 30–100 px) this can be 0.3–1.0; for a large nucleus (A ≈ 1000 px) it is ≈0.03. The reported ε≈0.05 is only an average. The same disks also label background pixels near boundaries, violating Eq. (12)/(16). The exact entropy at a nucleus pixel is H_binary(ε(x)), which is not proportional to P(CT) with a fixed constant; it is non-monotonic in ε when ε exceeds 0.5. Consequently, even a perfectly calibrated Bayesian network would not produce an entropy map proportional to P(CT) under the implemented sampling. The paper's Dice (0.71) and AUROC (0.944) only show that the entropy map separates nuclei from background; they do not validate the specific proportionality in Eq. (22). This is the load-bearing assumption of the theoretical contribution.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a weakly supervised nuclei instance detection pipeline. Starting from point annotations (one random pixel per nucleus plus a radius-3 disk), a Bayesian segmentation network is trained to output per-pixel predictive entropy. The paper argues in Appendix A that this entropy is proportional to the true nucleus probability with constant of proportionality -ε log ε, and uses this to justify converting the entropy map into a binary nucleus prior. The prior is then converted into instance masks through adaptive thresholding, Voronoi-edge subtraction, watershed, and filtering against the point annotations. These pseudo-masks are used to train Mask-RCNN, which achieves bounding-box mAP50 of 0.724 on PanNuke, compared to 0.834 when trained on full ground-truth masks, using about 5% of the pixel labels.","tokens_in":12388,"tokens_out":7897,"duration_ms":78196,"significance":"The paper's strength is the clean, transparent pipeline and the empirical demonstration that a weakly supervised model can approach full-supervision detection performance. The ablation studies (point radius, label percentage, position noise) are useful. The theoretical derivation, if made correct, would be a valuable contribution. However, as written, the derivation does not apply to the actual labeling procedure, and the key assumption that the network estimates the label-distribution entropy is not directly verified. These issues must be resolved before the theoretical claims can be accepted.","major_comments":[{"comment":"The derivation assumes a constant per-pixel label probability ε and zero probability of labeling background pixels as nucleus. The procedure in Section 3.2 samples one random point per nucleus and adds a radius-3 disk, so the conditional label probability at a true-nucleus pixel is ε(x) = area(B(x,3) ∩ nucleus)/area(nucleus), which depends on nucleus size and pixel location; for small nuclei this can be substantially larger than the reported average ε ≈ 0.05. The disk also spills over nucleus boundaries, labeling some background pixels as nucleus, violating Eq. (12)/(16). As a result, the entropy of the input label distribution is not of the form H ≈ -ε log ε P(CT) with a fixed constant, and Eq. (22) is not established for the implemented procedure. The theoretical justification of the bootstrap step therefore needs to be revised: either the labeling procedure should be changed to satisfy the assumptions (e.g., independent pixel-wise labeling with probability ε), or the derivation should be reworked for the actual variable-ε procedure, with appropriate approximation guarantees.","section":"Appendix A, Eqs. (12), (13), (16)"},{"comment":"The assumption that the Bayesian network is 'a good estimator of the entropy over the input label distribution' is not directly tested. The reported peak Dice (0.71) and AUROC (0.944) demonstrate that the entropy map separates nuclei from background, but they do not establish that the network's predictive entropy matches the analytical label-distribution entropy, which is what Eq. (22) requires. I recommend a direct comparison on a validation subset: compute the label-distribution entropy analytically from the sampling procedure and the ground-truth masks, and plot it against the network's predictive entropy (e.g., a binned scatter plot with correlation coefficient). This would provide direct evidence for the proportionality claim.","section":"Section 3.2"}],"minor_comments":[{"comment":"The two limit statements are incorrect as written: the ratio (εx log ε)/(εx log x) diverges as ε→0 (the denominator is a fixed negative constant for x<1), and the ratio (1-εx)log(1-εx)/(εx log x) tends to a constant, not 0. The final approximation H ≈ -ε log ε P(CT) is still correct, but the proof needs to be corrected to compare the terms properly.","section":"Appendix A, Eqs. (20)-(21)"},{"comment":"There are many typographical errors, including 'cell deection' (Section 1), 'Ee are only able' (Section 1), 'crefined' and 'crefer' (Section 3.2), and inconsistent spelling ('V oiculescu' in the author list). The manuscript would benefit from careful proofreading.","section":"Throughout"},{"comment":"The mapping from the radius-3 neighborhood to ε ≈ 0.05 is not explained; please state how this number is computed (e.g., average disk area divided by average nucleus area) and report the variance across nuclei.","section":"Section 3.2"},{"comment":"The paper does not compare against any existing weakly supervised cell/nuclei detection methods; a brief quantitative or qualitative comparison with one or two representative methods (even if on a different dataset) would help position the contribution.","section":"Section 4.2/4.3"},{"comment":"The 'peak Dice' is the maximum Dice over thresholds; this should be defined explicitly to avoid confusion with the Dice at a fixed threshold.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The theoretical flaw in Appendix A is significant because the paper explicitly advertises the derivation as a contribution. However, the empirical results are interesting and the method is simple and reproducible, so I believe the paper is worth major revision rather than rejection. The authors should also be encouraged to compare with existing weakly supervised methods and to add error bars or statistical significance to their tables."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth a look if you work on weakly supervised histopathology. It proposes a pipeline: train a Bayesian segmentation network on point labels, use the predictive entropy as a nucleus-likelihood map, convert it to instance masks via Voronoi edges, adaptive thresholding, and watershed, then train Mask-RCNN on those masks. The headline result is mAP50 0.724 on PanNuke with 5% of the pixel labels versus 0.834 with full masks.\n\nWhat is genuinely new is the idea that, under a sparse random-point labeling process, the entropy of a well-calibrated Bayesian network should be approximately proportional to the true foreground probability. Appendix A derives this cleanly. The empirical Dice (0.71) and AUROC (0.944) for the entropy map are solid, and the ablations on radius, label fraction, and position noise are useful. The paper is also honest that mAP75 drops sharply (0.30 vs 0.604) and that no comparison with other weakly supervised methods is provided.\n\nThe soft spot is that the theory and implementation do not quite line up. Appendix A assumes a constant per-pixel label probability epsilon (Eq. 13) and zero probability of labeling background (Eq. 16). But Section 3.2 samples one random pixel per nucleus and expands a radius-3 disk. That makes the conditional label probability for a nucleus pixel depend on nucleus size and the pixel's distance from the boundary; for small nuclei it can approach 1, and the disk will label some true-background pixels near boundaries. So Eq. (22) is not strictly established by the stated sampling procedure. A perfectly calibrated Bayesian network would still produce an entropy map that varies with epsilon(x), not a clean constant times P(CT). This does not sink the empirical method, since the entropy map clearly separates nuclei in practice, but the theoretical contribution overclaims. The authors should either model the actual epsilon(x) distribution or present the proportionality as a heuristic with empirical support.\n\nThe lack of comparison to other weakly supervised methods and the absence of code are also real weaknesses. Still, the method is simple, transparent, and the ablations are more thorough than typical.\n\nThis deserves a serious referee, but the theory section needs revision before acceptance.","headline":"Honest weak-supervision pipeline with a theory that oversimplifies the actual sampling—worth reviewing, but the theoretical claim needs tightening.","tokens_in":12907,"tokens_out":3248,"would_cite":true,"duration_ms":32872,"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":"One random point per nucleus lets a Bayesian network's entropy map stand in for a full segmentation mask, cutting labels 95% while keeping most detection accuracy.","keywords":["entropy bootstrapping","weakly supervised segmentation","nuclei detection","point annotations","Bayesian deep learning","instance segmentation","watershed","PanNuke"],"falsifier":"Train the same Bayesian segmentation network on point labels that violate the sparsification model, for example labels biased toward nuclear centers or with a small fraction of background pixels deliberately labeled as nucleus, and compare the thresholded entropy map directly with ground-truth masks. The paper's derivation predicts the H ∝ P(CT) proportionality should degrade under such violations, so a large drop in peak Dice or AUROC would confirm that the identity is doing the work; if performance is unchanged, the entropy map would be learning something more general than the claimed mechanism.","tokens_in":11883,"feed_emoji":"🔬","tokens_out":6895,"duration_ms":67483,"temperature":0.7,"pith_summary":"This paper tries to establish that a Bayesian segmentation network trained on just one point label per nucleus produces an entropy map that is proportional to the true probability that a pixel is nucleus, and that this map can be converted into instance masks good enough to train a standard detector. The appeal is practical: annotating a single dot per nucleus instead of drawing full contours cuts labeled pixels by 95%. On PanNuke, training Mask-RCNN on the bootstrapped masks reaches a bounding-box mAP50 of 0.724, against 0.834 when trained on full ground-truth masks. The paper also shows the entropy map degrades gracefully when point labels are fewer or slightly misplaced.","feed_headline":"One dot per nucleus keeps 87% of full-mask detection accuracy","feed_subtitle":"Bayesian entropy turns point labels into masks that train Mask-RCNN to mAP50 0.724 on PanNuke.","key_machinery":"The central object is the proportionality identity H ∝ P(CT) (Eq. 22), derived in Appendix A from the assumption that foreground point labels occur with small probability epsilon and background labels are exact. This identity is what lets an entropy map double as a segmentation prior. The rest of the pipeline is a deterministic conversion: subtract Voronoi edges of the point labels, apply adaptive thresholding, watershed within regions of interest, then keep only masks that coincide with the original point labels. Mask-RCNN is the final refinement stage that turns the rough watershed instances into the reported detection outputs.","core_discovery":"Under the paper's sparsification model, each nucleus is represented by a small randomly placed labeled neighborhood of probability epsilon, and background pixels are assumed labeled correctly. The paper derives that the predictive entropy of a Bayesian segmentation network trained on these labels behaves as H = -epsilon log(epsilon) P(CT) for small epsilon, so the entropy map is a constant multiple of the true nucleus probability. This identity turns the network's uncertainty output into a binary nuclei prior without any contour supervision. The prior is then thresholded and separated with Voronoi edges, adaptive thresholding, and watershed, and the resulting instance masks are used as training targets for Mask-RCNN. The paper reports that this pipeline localises nuclei nearly as well as full-mask training at the IoU 0.5 level, with the main shortfall appearing at stricter IoU thresholds.","pith_inferences":["Beyond the paper: the proportionality identity suggests the bootstrap is not specific to nuclei; any point-labeled structure detection task where a sparse foreground label model holds could reuse the same entropy-to-prior conversion.","Beyond the paper: the derivation exposes an annotation-budget trade-off the paper only samples: epsilon must be small enough for the -epsilon log(epsilon) term to dominate the entropy, but large enough for the network to learn from the sparse labels; a systematic sweep over epsilon could locate the optimal operating point for other datasets.","Beyond the paper: using the entropy map directly as a soft target or as an uncertainty weighting during Mask-RCNN training could preserve more information than the hard thresholded masks, potentially narrowing the gap to the full-mask baseline.","Beyond the paper: the loss of calibrated uncertainty at the Mask-RCNN stage, acknowledged in the paper, suggests that replacing the final detector with an uncertainty-aware head would restore the transparent error estimates the earlier stages provide."],"forward_implications":["With roughly 5% of the pixel labels (epsilon ≈ 0.05), the bootstrapped masks train Mask-RCNN to a bounding-box mAP50 of 0.724 on PanNuke, compared with 0.834 from full masks; segmentation mAP50 is 0.632 versus 0.826.","The entropy map remains a useful discriminator under weaker labels: halving the number of point annotations lowers mean AUROC only from 0.944 to 0.904, and adding up to 5 pixels of position error lowers it to 0.926.","The deterministic watershed output alone detects nuclei at a rate of 0.61 at IoU 0.5, so most of the detection signal is present before the learned refinement step.","Because the first stages are statistically grounded rather than learned, the pipeline is transparent up to the Mask-RCNN step, where calibrated uncertainty is lost."],"supporting_citations":[{"why":"Supplies the Functional Variational Inference Bayesian segmentation network whose predictive entropy is the bootstrap signal.","marker":"[1]"},{"why":"Provides the PanNuke dataset, including the ground-truth masks from which point labels are sampled and against which results are measured.","marker":"[5]"},{"why":"Supplies Mask-RCNN, the off-the-shelf instance segmentation model trained on the bootstrapped watershed masks.","marker":"[8]"}],"fun_headline_variants":["Entropy bootstrapping: point labels to full masks in one step","One point per nucleus: entropy does the rest","Point-only labels? Entropy bootstrapping fills the gaps","95% label cut: entropy gives near-full nuclei detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Bayesian segmentation network's predictive entropy faithfully estimates the entropy of the input label distribution; if that estimate drifts, the claimed proportionality between entropy and nucleus probability collapses.","fun_headline_variants_meta":{"raw":{"variants":["Entropy bootstrapping: point labels to full masks in one step","One point per nucleus: entropy does the rest","Point-only labels? Entropy bootstrapping fills the gaps","95% label cut: entropy gives near-full nuclei detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002565,"raw_usage":{"total_tokens":9761,"prompt_tokens":821,"completion_tokens":8940,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":437,"completion_tokens_details":{"reasoning_tokens":8871}},"tokens_in":437,"tokens_out":8940,"duration_ms":63002,"temperature":1.0,"reasoning_tokens":8871,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:17:23.764410+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same Bayesian segmentation network on point labels that violate the sparsification model, for example labels biased toward nuclear centers or with a small fraction of background pixels deliberately labeled as nucleus, and compare the thresholded entropy map directly with ground-truth masks. The paper's derivation predicts the H ∝ P(CT) proportionality should degrade under such violations, so a large drop in peak Dice or AUROC would confirm that the identity is doing the work; if performance is unchanged, the entropy map would be learning something more general than the claimed mechanism.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Functional Variational Inference Bayesian segmentation network whose predictive entropy is the bootstrap signal."},{"cited_title":"Pannuke: an open pan- cancer histology dataset for nuclei instance segmentation and classification","cited_arxiv_id":null,"evidence_quote":"Provides the PanNuke dataset, including the ground-truth masks from which point labels are sampled and against which results are measured."}],"review_version":1}