{"id":"d2c607a1-227e-4365-80e9-256649c8d48e","arxiv_id":"2505.04150","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"OSLSP uses a similarity-proportion loss computed from pairs of weakly labeled bags to update a feature extractor for ordinal classes, and it reports higher accuracy than DINO pre-trained or fine-tuned baselines on a 31-WSI skeletal muscle recovery dataset.","lead":"This paper introduces a training method that lets an image classifier learn muscle-cell recovery stages from whole-slide images using only rough per-day proportions instead of cell-by-cell labels. A reader might care because it is a step toward automated, objective scoring of tissue regeneration from standard histology slides.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Attribution of OSLSP's reported gain to the similarity proportion loss is not established: Table 2 lacks an ablation that removes L_SimProp, and the 46.0% vs 44.4% margin over pre-trained DINO rests on 5 test WSIs with no error bars.","rationale":"The reader's verdict is CONDITIONAL, and I agree that the method is plausible: the loss is clearly motivated, the Gaussian-expansion trick makes the histogram differentiable, and the comparison to DINO is a reasonable first step. The most load-bearing weakness is attribution: the proposed similarity proportion loss is never isolated from the rest of the training pipeline. The reader's weakest_assumption about the hand-set linear similarity in Eq. 5 and the rough per-date proportions in Eq. 6 is real, but it is testable by the same ablation design: a wrong-ordering negative control distinguishes 'the ordinal prior is working' from 'the loss is just fine-tuning the backbone.' I credit the paper for a concrete medical problem and for explicitly acknowledging the intact/ghost fiber confusion in the discussion. However, there is no code, no data release, and no statistical inference, so the empirical claim should be re-evaluated under the specified ablation before acceptance. Because the concern is about missing evidence rather than an identified falsehood or internal inconsistency, the appropriate verdict remains CONDITIONAL, which is unchanged from the reader's verdict.","tokens_in":7641,"tokens_out":11333,"duration_ms":121441,"concrete_test":"Run a controlled ablation on the same 5-WSI test split: (A) OSLSP as reported; (B) the identical pipeline with L_SimProp removed, keeping L_prop and last-block fine-tuning; and (C) the identical pipeline with L_SimProp but a permuted class order so the ordinal prior is wrong. Report per-WSI accuracy, macro F1, and a paired bootstrap or McNemar test across the 5 WSIs. If B matches A within the bootstrap interval, or if C also beats pre-trained DINO, then neither the similarity proportion loss nor the ordinal prior is load-bearing; if A significantly beats B, C, and DINO, the central claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two parts: that OSLSP can update the feature extractor under LLP, and that it outperforms baselines. The experimental support for both is weaker than Table 2 suggests. OSLSP is trained with L_SimProp plus L_prop while fine-tuning only the last block; the only non-OSLSP feature-extractor condition is a fully fine-tuned DINO that collapses to 20.967% accuracy. There is no condition with the same last-block fine-tuning and L_prop alone, nor a condition with L_SimProp but a scrambled class order. Thus the +1.56 percentage-point gain over pre-trained DINO could be caused by the fine-tuning procedure, augmentation, or the trained head rather than by the ordinal similarity proportion loss. The fine-tuned baseline's collapse also suggests the comparison is not an apples-to-apples optimization setup. With only 5 test WSIs and no confidence intervals, the accuracy gap is within plausible sampling noise, so the headline 'outperforms' is not statistically supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes OSLSP, an LLP method for classifying five ordinal skeletal muscle recovery stages from cell images extracted from WSIs. The method computes a differentiable similarity-proportion loss between pairs of bags, in which the ground-truth similarity distribution is obtained by combining a hand-set class order with per-date class proportions, and the predicted distribution is obtained from cosine similarities of instance features. OSLSP updates the last block of a pre-trained ViT-B/8 backbone together with the classifier, and the authors report that it achieves 46.005% micro accuracy and 0.425 F1 on five test WSIs, compared with 44.442% accuracy and 0.377 F1 for the pre-trained DINO baseline and 20.967% accuracy and 0.217 F1 for fine-tuned DINO.","tokens_in":7881,"tokens_out":3896,"duration_ms":37429,"significance":"If the reported results were statistically supported, the contribution would be useful for the medical-imaging LLP community: the paper proposes a principled, differentiable way to use ordinal class structure under label-proportion supervision, and the idea of updating a feature extractor through the similarity-proportion loss addresses a real limitation of previous LLP methods for WSIs. The paper deserves credit for anchoring the loss in a concrete biological ordering of recovery stages and for using Gaussian expansion to make the histogram-based loss differentiable. However, the current evidence base is too thin to support the headline claim that OSLSP outperforms DINO-based models: there are only five test WSIs, no error bars or significance tests, no ablation that removes the similarity-proportion term, and the fine-tuned baseline uses a different training recipe.","major_comments":[{"comment":"The evaluation rests on only five test WSIs (one per day), and the reported metrics are point estimates without confidence intervals, error bars, or significance tests. The accuracy gap between OSLSP (46.005%) and pre-trained DINO (44.442%) is 1.56 percentage points; with only five independent test slides, this gap is well within plausible sampling variability. Please report per-slide scores, bootstrap confidence intervals, or a paired significance test, and state the number of cells per slide and per class.","section":"Section 4, Table 2"},{"comment":"There is no ablation that isolates the contribution of L_SimProp. The only condition that updates a feature extractor is a fully fine-tuned DINO using a different training recipe (the official DINO pipeline), and it collapses to 20.967% accuracy. To attribute the gain to the ordinal similarity-proportion loss, add at least a last-block fine-tuning condition with Lprop only (removing L_SimProp) and an OSLSP variant trained with a scrambled or reversed class order. Without these controls, the 1.56-point gain over the pre-trained baseline could be caused by the fine-tuning procedure, data augmentation, or the trained head rather than by the proposed loss.","section":"Section 3.2 and Table 2"},{"comment":"The ground-truth similarity distribution P is determined by the hand-set linear similarity sim(k,k') = 1 - |k-k'|/(K-1) and by rough manual class proportions p_d; both are treated as fixed prior knowledge. If the linear spacing misrepresents the true morphological distances, or the rough proportions are noisy, the loss will systematically drive the feature extractor toward an incorrect target. Since the RMSE evaluation metric uses the same linear class order, the ordinal evaluation shares that assumption. Please provide a sensitivity analysis (e.g., vary the similarity matrix, sigma, and the proportion estimates) or validate the ordinal prior against an independent annotation.","section":"Section 3.1, Eqs. (5)-(6)"}],"minor_comments":[{"comment":"The text says 'all instance pairs' but the summation in Eq. (4) runs over n = 1,...,N, which suggests only N paired instances rather than all N^2 pairs; please clarify which pairing is used, as this changes the effective training distribution.","section":"Section 3.1, Eq. (4)"},{"comment":"The phrase 'inter-atomic distances' appears to be a typo for 'instance similarities' or 'pairwise distances'.","section":"Section 3.1"},{"comment":"The paper does not state the number of instances (cells) per WSI or per class; reporting these numbers is needed to interpret the macro precision, recall, and F1 scores, which can be very volatile with only five test slides.","section":"Section 4"},{"comment":"A confusion matrix would help quantify the acknowledged confusion between intact MF and ghost fiber and would make the ordinal error structure visible; the current RMSE alone does not reveal which class pairs are confused.","section":"Section 4, Table 2 and Section 5"},{"comment":"The claim that the model 'outperforms' large-scale pre-trained and fine-tuning models should be tempered to 'on the reported dataset' unless statistical support is added.","section":"Abstract and Section 5"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"OSLSP is a genuinely new loss function for learning from label proportions with ordinal classes, and the problem it addresses is real. The similarity proportion loss compares two bags by their pairwise cosine similarity histograms and matches them to a ground-truth distribution derived from the class proportions via an ordinal similarity prior. That is a clever construction, and it is not in the cited prior work. The paper also updates the feature extractor rather than just the head, which is a known gap in LLP. The math at the equation level is coherent, and the authors are honest about the confusion between intact and ghost fibers.\n\nThe empirical support, though, is too weak to back the headline claim. Five test WSIs, no error bars, no significance tests. The gain over pre-trained DINO is 1.6 percentage points, which could easily be sampling noise. More importantly, there is no ablation that removes L_SimProp while keeping the same fine-tuning recipe. The only non-OSLSP feature-extractor condition is a fully fine-tuned DINO that collapses to 21% accuracy, and that is a different training setup. So the attribution of the gain to the similarity proportion loss is not established. The RMSE metric shares the same linear ordering used to construct the training prior, which is a mild form of circularity; the accuracy numbers are cleaner.\n\nThe paper is worth a serious referee. The method is novel enough, and the limitation statements in the Discussion show the authors understand their own weaknesses. But a referee would need to push for a larger test set, confidence intervals, a proper ablation, and a baseline that only removes the similarity proportion loss. As it stands, the paper should not be accepted without those changes. If I were the editor, I would send it to review with a clear request for revision.","headline":"A genuinely new ordinal LLP loss that deserves a serious referee, but the reported gain rests on five test WSIs and no ablation isolating the loss.","tokens_in":8405,"tokens_out":2148,"would_cite":false,"duration_ms":21012,"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 similarity proportion loss trains an ordinal classifier of muscle recovery stages under weak supervision, beating large pre-trained and fine-tuned DINO baselines.","keywords":["learning from label proportions","ordinal classification","whole slide images","skeletal muscle regeneration","similarity proportion loss","weakly supervised learning","feature extractor fine-tuning","cell morphology"],"falsifier":"Compute the class-pair confusion matrix on held-out WSIs for OSLSP and for the pre-trained DINO baseline; the paper's own Fig. 2 suggests OSLSP confuses intact myofibers and ghost fibers more often. If replacing the linear similarity with a squared or exponential spacing, or with a dissimilarity matrix estimated from a few expert-annotated regions, does not change test accuracy and F1, then the linear-ordering prior is not what drives the result.","tokens_in":7442,"feed_emoji":"🔬","tokens_out":6070,"duration_ms":58390,"temperature":0.7,"pith_summary":"To classify cells in whole-slide images of regenerating skeletal muscle without per-cell labels, the paper proposes OSLSP, an LLP method that uses the date after cardiotoxin injection as weak supervision. The method derives a similarity proportion loss from pairs of bags, comparing the histogram of cosine similarities between instances in two bags with a ground-truth similarity distribution built from the two bags' class proportions and a hand-set linear ordering of the five recovery stages. Because the histogram is computed with Gaussian expansion, the loss is differentiable and can update the feature extractor, not just the classification head. On the authors' test set, OSLSP reaches 46.005% micro accuracy and F1 0.425, compared with 44.442%/0.377 for the pre-trained DINO model and 20.967%/0.217 for a fine-tuned DINO model. The main assumption is that the linear similarity between stages, $\\mathrm{sim}(k,k') = 1 - |k-k'|/(K-1)$, correctly encodes the biological order and spacing of recovery.","feed_headline":"A proportion-only loss beats DINO at muscle recovery staging","feed_subtitle":"OSLSP learns cell-stage order from rough per-day proportions, reaching 46.0% accuracy on whole-slide images.","key_machinery":"The central object is the similarity proportion loss $L_{\\mathrm{SimProp}} = D_{\\mathrm{KL}}(\\hat{P} \\parallel P)$, built by grouping same-date instances into bags, computing a differentiable histogram of pairwise cosine similarities between two bags using Gaussian expansion with $\\sigma = 0.1$, and forming the ground-truth histogram as $P(\\mathrm{sim}(k,k')) = p_k p'_{k'} + p_{k'} p'_k$ (with the diagonal case $k = k'$ using $p_k p'_k$) from the class proportions and the linear class similarity $\\mathrm{sim}(k,k') = 1 - |k-k'|/(K-1)$. This loss is differentiable, so gradients reach the backbone, which standard LLP losses cannot do, and the same construction injects the ordinal prior into training.","core_discovery":"OSLSP's central claim is that label proportions can supervise both the classifier and the feature extractor in an ordinal setting. For two bags with class proportion vectors $\\mathbf{p}_d$ and $\\mathbf{p}_{d'}$, the ground-truth similarity distribution is formed by pairing classes and weighting each class-pair by the product of its proportions, with class similarity $\\mathrm{sim}(k,k') = 1 - |k-k'|/(K-1)$ encoding the biological life cycle of myofibers (intact myofiber, ghost fiber, myoblast, myotube, recovered myofiber). The predicted distribution is a differentiable histogram of scaled cosine similarities between instance features from the two bags, and the KL divergence between the two distributions is the similarity proportion loss. Training with this loss plus a standard proportion loss for the head adapts a ViT-B/8 backbone to muscle morphology while respecting stage order. The paper reports that this combination beats a large pre-trained DINO feature extractor and a fine-tuned DINO model on accuracy, F1, and RMSE, while conceding that intact myofibers and ghost fibers are confused more often than in the pre-trained model because their pre-set similarity is high.","pith_inferences":["The linear similarity $\\mathrm{sim}(k,k') = 1 - |k-k'|/(K-1)$ assumes equal spacing between consecutive stages; a testable extension is to replace it with an exponential or learned kernel, or with a dissimilarity matrix estimated from a few expert-annotated regions, and check whether accuracy and F1 improve.","Since the loss is agnostic to how bags are formed, it could be applied to other grouping variables such as subject, batch, or stain in weakly supervised pathology tasks where an ordinal progression is known, not just to days after injection.","The reported gain over the pre-trained DINO baseline is modest on a single five-WSI test set, so the more durable contribution may be the demonstration that backbone fine-tuning under LLP is feasible, rather than the exact accuracy difference over one dataset.","Combining the similarity proportion loss with a consistency regularizer similar in spirit to LLP-VAT could further stabilize the feature extractor, especially in low-data regimes where the rough manual proportions are noisy."],"forward_implications":["LLP on whole-slide images no longer requires a frozen pre-trained feature extractor; the backbone can be adapted under weak proportion supervision when class order is meaningful.","The two-bag similarity distribution construction can be applied to any ordinal or structured label space, such as disease severity scores or developmental time points, when rough proportions per group are available.","Because the loss uses only pairwise cosine similarities, it avoids the pseudo-label noise that instance-level MIL methods such as IIB-MIL introduce, while still updating the backbone.","The reported RMSE improvement over the pre-trained DINO baseline (2.152 vs. 2.431) indicates that OSLSP makes fewer far-off ordinal errors, which matters for assessing muscle regeneration where confusing adjacent stages has different consequences than confusing distant ones.","The confusion between intact myofibers and ghost fibers, which the paper attributes to the high pre-set similarity between those classes, shows that the ordinal prior directly shapes the model's error profile rather than merely improving overall accuracy."],"supporting_citations":[{"why":"Supplies the pre-trained ViT-B/8 backbone and the two baselines (frozen and fine-tuned) that OSLSP must beat.","marker":"[2]"},{"why":"Supplies the standard LLP bag-making procedure that groups same-date instances into a bag.","marker":"[1]"},{"why":"Provides the Gaussian expansion that makes histogram counts differentiable, enabling gradient flow to the feature extractor.","marker":"[13]"},{"why":"Provides the class-weighting analogy in supervised learning that motivates injecting ordinal information through class similarities.","marker":"[5]"},{"why":"Represents a previous MIL method that updates both backbone and head using pseudo-labels, which OSLSP contrasts with proportion-only supervision.","marker":"[11]"},{"why":"Provides the cell segmentation model used to crop instance images from whole-slide images.","marker":"[16]"},{"why":"Exemplifies LLP applied to whole-slide images with head-only learning, the setting OSLSP extends.","marker":"[19]"},{"why":"Shows LLP applied to whole-slide image segmentation while treating classes nominally, motivating OSLSP's ordinal treatment.","marker":"[8]"},{"why":"Motivates the need for region-specific evaluation of cell condition in regenerating muscle tissue.","marker":"[4]"}],"fun_headline_variants":["OSLSP loss outperforms DINO on muscle recovery staging","Proportion loss with ordinal scale wins on muscle staging","Learn muscle recovery stages from label proportions alone","Ordinal label proportions beat fine-tuned DINO at muscle staging"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the hand-set linear class similarity $\\mathrm{sim}(k,k') = 1 - |k-k'|/(K-1)$ correctly encodes the biological order and spacing of the five recovery stages, and that the roughly estimated per-date class proportions are accurate enough to serve as ground truth in the supervision signal.","fun_headline_variants_meta":{"raw":{"variants":["OSLSP loss outperforms DINO on muscle recovery staging","Proportion loss with ordinal scale wins on muscle staging","Learn muscle recovery stages from label proportions alone","Ordinal label proportions beat fine-tuned DINO at muscle staging"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000172,"raw_usage":{"total_tokens":1311,"prompt_tokens":1014,"completion_tokens":297,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":231}},"tokens_in":630,"tokens_out":297,"duration_ms":3575,"temperature":1.0,"reasoning_tokens":231,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:35:57.758139+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the class-pair confusion matrix on held-out WSIs for OSLSP and for the pre-trained DINO baseline; the paper's own Fig. 2 suggests OSLSP confuses intact myofibers and ghost fibers more often. If replacing the linear similarity with a squared or exponential spacing, or with a dissimilarity matrix estimated from a few expert-annotated regions, does not change test accuracy and F1, then the linear-ordering prior is not what drives the result.","supporting_citations":[{"cited_title":"LLP-Bench: A Large Scale Tabular Benchmark for Learning from Label Proportions","cited_arxiv_id":"2310.10096","evidence_quote":"Supplies the standard LLP bag-making procedure that groups same-date instances into a bag."},{"cited_title":"In: Medical Image Computing and Computer Assisted Intervention (MICCAI)","cited_arxiv_id":null,"evidence_quote":"Represents a previous MIL method that updates both backbone and head using pseudo-labels, which OSLSP contrasts with proportion-only supervision."},{"cited_title":"Skeletal Muscle9 (2019)","cited_arxiv_id":null,"evidence_quote":"Motivates the need for region-specific evaluation of cell condition in regenerating muscle tissue."}],"review_version":1}