{"id":"54072580-c211-42f5-af97-87332f85b9fd","arxiv_id":"2412.02976","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SADA combines stain-based augmentation, local alignment, and anchor-averaged supervised contrastive learning to improve imbalanced blood cell classification across domains.","lead":"SADA is a blood cell classification method that re-stains cell images with colors from other datasets and trains the model to ignore stain differences while learning class features. On four public datasets and one hospital dataset it reports higher F1 scores than several existing methods, suggesting a practical way to make blood cell AI more robust across labs.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Hyperparameter selection appears to use the held-out test domains, potentially inflating the reported gains.","rationale":"The reader's weakest-assumption choice (SNMF decomposition validity) is reasonable but, in my view, secondary. Even if the stain/structure separation is imperfect, SADA could still empirically outperform baselines because the augmentation, alignment, and contrastive losses may still provide useful regularization. The reported quantitative superiority is the central claim, and that claim is directly threatened by the apparent use of test data for hyperparameter tuning. The paper's Section IV-D explicitly says the tuning was carried out under the leave-one-out framework of Section IV-A, meaning the target domains were included in the selection of k and beta. No source-only validation set is described, and the baselines are not given the same target-aware tuning, so the comparison is potentially unfair. A concrete re-run with proper source-domain validation would settle whether the state-of-the-art margin persists. Since the paper otherwise has merit and no clear internal contradiction, the correct verdict remains CONDITIONAL, but with attention shifted to evaluation integrity rather than the SNMF mechanism.","tokens_in":12968,"tokens_out":5676,"duration_ms":53814,"concrete_test":"Repeat the leave-one-out experiments for SADA and the baselines, but select k and beta only on a validation split drawn from the source domains (e.g., 10% of each training dataset), never using the held-out target domain for hyperparameter choice. Then evaluate on the target. If SADA's average F1-micro and F1-macro drop by more than 2 points from the reported 84.4% / 68.6%, or if the margin over BoDA falls below statistical significance (e.g., overlapping error bars), the reported advantage is inflated by target-aware tuning.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is state-of-the-art performance, but the evaluation protocol appears to leak target-domain information into hyperparameter selection. In Section IV-D, the authors state that they tune the number of clusters k and the local-alignment weight beta by measuring F1-micro and F1-macro under the framework of Section IV-A, which is the leave-one-out protocol that uses each public dataset as the held-out target. This implies the same folds used to report final accuracy were also used to select the best hyperparameters, so the numbers in Table I (84.4% F1-micro, 68.6% F1-macro, exceeding BoDA by 4.4% and 3.9%) may be optimistically biased. No separate validation split is mentioned, and the baselines are not described as undergoing the same target-aware tuning. If the hyperparameters were instead chosen on source-domain validation data, the margin over the second-best method could shrink or vanish, undermining the state-of-the-art claim. This is more direct and load-bearing than the stain-decomposition concern because it affects the measured outcome itself, not just the proposed mechanism.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SADA, a two-stage method for blood cell classification under domain shift and class imbalance. In Stage 1, the method performs stain-based augmentation by decomposing images into stain color matrices and density maps via Beer-Lambert and sparse non-negative matrix factorization, clustering the color bases, and re-staining density maps with color bases from other clusters. A pixel-wise local alignment loss and a domain-invariant supervised contrastive loss are then applied to the raw and augmented samples. In Stage 2, a classifier is trained with standard cross-entropy. The method is evaluated with a leave-one-out protocol on four public blood cell datasets and on a private hospital dataset, reporting F1-micro and F1-macro. The central claim is state-of-the-art performance, with average F1-micro of 84.4% and F1-macro of 68.6% on the public datasets, exceeding the second-best method BoDA by 4.4% and 3.9%, respectively.","tokens_in":13131,"tokens_out":4473,"duration_ms":48357,"significance":"If the reported results are valid, SADA would be a useful practical contribution to hematological image analysis, where both domain shift and data imbalance are common. The paper has several strengths: it uses a standard and challenging cross-dataset evaluation protocol, reports results over three trials, includes an ablation study, and provides a source code URL. The proposed components are simple and additive, which makes the method easy to reproduce. However, the validity of the headline claim depends on two load-bearing issues: hyperparameter selection appears to use the same held-out target domains as the final evaluation, and the stain decomposition assumption underlying the augmentation is not independently validated. These issues, if left unaddressed, make the reported margins over the baselines difficult to interpret as honest generalization performance.","major_comments":[{"comment":"Hyperparameter selection in Section IV-D is performed using the same leave-one-out protocol that produces the final results in Table I. The text states that k and beta were tuned by measuring F1-micro and F1-macro 'under the framework outlined in Section IV-A,' and Section IV-A is the leave-one-out evaluation protocol. Therefore the values k=3 and beta=0.1 were selected using the held-out target domains, and the reported averages of 84.4% F1-micro and 68.6% F1-macro are the result of target-aware selection rather than independent prediction. This selection bias directly affects the central state-of-the-art claim. The paper should either use a nested or source-validation protocol for hyperparameter selection, or report results for a fixed, pre-specified configuration, and should clarify whether the baseline methods in Tables I and II were given equivalent target-aware tuning.","section":"Section IV-D, Table I"},{"comment":"The stain-based augmentation method assumes that the SNMF decomposition in Eq. (1) separates stain appearance W from cell structure H, so that recoloring H with W from another k-means cluster preserves morphology while simulating domain shift. This assumption is load-bearing because the local alignment loss and the domain-invariant supervised contrastive loss both train on the reconstructed samples. The paper provides no independent validation that the decomposition preserves labels or morphological structure. Since the ablation study in Table III does not include a variant without the stain-based augmentation module, the contribution of this assumption is not isolated. A concrete validation would be to have the generated samples checked by a human expert, or to measure classification-consistent morphology before and after re-staining, or to run an ablation that removes SBA while keeping the other losses.","section":"Section III-A, Eqs. (1)-(3)"},{"comment":"The paper repeatedly states that decoupling training into two stages alleviates data imbalance, but Stage 2 trains the classifier with plain cross-entropy, Eq. (8), and no class-balanced sampling, re-weighting, or other imbalance-aware mechanism is described. The only imbalance-aware components are the supervised contrastive loss, which considers all positives regardless of class frequency, and the macro-F1 metric. The reported improvement in macro-F1 is therefore not attributable to the two-stage schedule unless an ablation or analysis isolates it. The authors should either add a class-balanced classifier stage or provide a dedicated experiment showing that the two-stage procedure, beyond the proposed losses, is responsible for the macro-F1 gains.","section":"Section III-D, Eq. (8), Table I"}],"minor_comments":[{"comment":"In the sentence introducing density maps and stain color matrices, the text says 'density maps Hs (i.e. stain color) and stain color matrices Ws (i.e. structure)', which reverses the definitions given in the same section and in Eq. (1). Density maps should be associated with structure and stain color matrices with appearance.","section":"Section III-A"},{"comment":"The summation index k in the denominator of Eq. (6) conflicts with the hyperparameter k used for the number of stain-color clusters in Section III-A. A different index, such as m or j, should be used for the summation.","section":"Section III-C, Eq. (6)"},{"comment":"Contribution 2 says the local alignment constraint makes 'the target sample and generated samples' consistent, but no target-domain samples are used during training. The intended meaning appears to be 'raw sample and generated samples,' and the wording should be corrected.","section":"Section I, contribution list"},{"comment":"Figure 7 reports only the hyperparameter tuning results averaged over the four public datasets, without per-dataset curves or error bars. It would be clearer to state how the optimal k and beta were selected from the plotted curves and whether the same values were used without further tuning for the SYSU3H external test in Table II.","section":"Section IV-D, Fig. 7"},{"comment":"The sentence 'Regarding Loss re-weighting addresses data imbalance by adjusting the loss contribution of each class' is grammatically incomplete; it should be rewritten, for example as 'Regarding loss re-weighting, it addresses data imbalance by adjusting the loss contribution of each class.'","section":"Section II-A"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant applied problem and the method is straightforward to implement, but the evaluation protocol currently leaks target-domain information into hyperparameter selection, which undermines the strength of the reported state-of-the-art claim. I would like the editor to ensure that the revised version reports results under a proper nested tuning protocol or with fixed hyperparameters justified on source validation data only, and that the baseline methods are given comparable treatment. The stain-decomposition validation request is also important because the augmentation step is central to the method's mechanism."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: SADA is a sensible combination of existing components—SNMF stain decomposition, k-means pseudo-domains, local alignment, and anchor-averaged supervised contrastive loss—and it appears to work. The gains over BoDA and other baselines are consistent across four public datasets and a private hospital set, and the external validation is a nice touch. But the reported numbers are not trustworthy as state-of-the-art claims because Section IV-D says k and beta were chosen by measuring F1 under the Section IV-A protocol, which is the leave-one-out evaluation itself. That means the same folds used to report final accuracy were used to pick hyperparameters. The baselines are not described as having the same luxury. This is a selection-bias problem, not a fatal one: the margins on some datasets are large enough that the method likely still helps, but the magnitude of the edge is probably smaller than claimed.\n\nThe main new content is the anchor averaging in the contrastive loss, which is a reasonable idea to reduce domain-specific influence on anchors. The rest is recombination. The paper doesn't compare against the closest stain-augmentation baselines (StainMixUp, RandStainNA), which is a real gap given the novelty claim rests on the augmentation design. The ablation table is also confusing: M3 is described as the full method with standard supervised contrastive loss, but the table omits LLA for M3. That's either a typo or a different configuration, and it should be fixed. No significance tests, and the code is claimed but not actually linked in the full text.\n\nThe stain-decomposition assumption—that W captures stain and H captures morphology—is not independently validated, but the empirical success suggests it's reasonable. I don't see a circularity problem at the equation level.\n\nOverall: this is a practical paper for medical imaging and domain generalization researchers. It's not a breakthrough, but the method is plausible, the evaluation is standard (modulo the hyperparameter issue), and the problem is real. With a proper nested validation and the missing baselines, the claims would be credible. I'd recommend a serious referee look at it, with major revision. For my own work, I'd cite it as an example of stain-aware augmentation for domain generalization.","headline":"Useful practical method for imbalanced blood cell classification, but the reported edge over baselines is likely inflated because hyperparameters were tuned on the held-out test domains.","tokens_in":13714,"tokens_out":3416,"would_cite":true,"duration_ms":31225,"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":"This paper claims that a two-stage, stain-aware training procedure can make blood-cell classifiers robust to both domain shift and class imbalance.","keywords":["domain generalization","data imbalance","blood cell classification","stain normalization","contrastive learning","Beer-Lambert decomposition","feature alignment","long-tailed recognition"],"falsifier":"Take a single blood-cell image, decompose it via the paper's SNMF procedure, re-stain its density map with stain matrices from the other k-means clusters, and compare a cell-morphology measurement such as nucleus segmentation overlap or cell boundary IoU between the original and re-stained versions; if morphology shifts systematically with the chosen stain matrix, the separation premise fails.","tokens_in":12743,"feed_emoji":"🩸","tokens_out":5550,"duration_ms":56530,"temperature":0.7,"pith_summary":"The paper introduces SADA, a blood-cell classification method built around the idea that stain color is the main domain-specific appearance factor while cell morphology is domain-invariant. It tries to establish that by decomposing blood-cell images into stain color and structure, recoloring the structure with other domains' stain profiles, and then aligning features between original and recolored views, a classifier can transfer to unseen hospital datasets and still handle rare cell classes. The reported evidence is a leave-one-out evaluation across four public blood-cell datasets, where SADA reaches average F1-micro of 84.4% and F1-macro of 68.6%, beating the second-best method by 4.4 and 3.9 points, plus a private hospital dataset where it scores 85.4% and 85.1%. A sympathetic reader would care because the method targets exactly the two obstacles that keep automated blood-cell counting out of clinical practice: instruments and staining protocols differ across sites, and rare cell types are easily crushed by common ones.","feed_headline":"Stain-aware blood-cell classifier beats domain shift by 4.4 points","feed_subtitle":"Re-staining cells with other labs' colors teaches the model shape features that transfer across unseen datasets.","key_machinery":"The load-bearing object is the Beer-Lambert optical-density factorization $V = W H$, obtained via sparse non-negative matrix factorization, which splits a blood-cell image into a stain color matrix $W$ and a density map $H$ that is assumed to carry the cell structure. The method clusters the $W$ matrices in each mini-batch, then re-stains a sample's density map with the $W$ from other clusters to create synthetic domain variants while leaving morphology intact. A pixel-wise local alignment loss and a contrastive loss whose anchors are averages of original and re-stained embeddings then drive the feature extractor to ignore the differences between these views.","core_discovery":"SADA's central claim is that domain-invariant features can be learned by exploiting the Beer-Lambert structure of stained microscopy images. The method first factorizes each image's optical density into a stain color matrix and a structure density map using sparse non-negative matrix factorization, clusters the stain colors in each mini-batch, and then generates synthetic domain-transformed samples by re-staining the original structure with stain colors from other clusters. A pixel-wise local alignment loss forces the backbone's feature maps to agree between original and re-stained views, and a domain-invariant supervised contrastive loss builds anchors by averaging the embeddings of all views of the same cell, diluting domain-specific content. The paper reports that this training, followed by a second stage of classifier training on frozen features, yields the highest F1-micro and F1-macro scores on each of the four public datasets under leave-one-out evaluation, and also on a held-out private clinical dataset, with the largest gains over prior methods on the datasets where stain variation is most visible.","pith_inferences":["The same re-staining operation could be applied at test time to normalize an incoming image into a canonical stain profile before classification, a post-deployment robustness measure the paper does not explore.","A factorial ablation separating the diversity benefit of the augmentation from the alignment benefit of the losses would clarify whether the 4.4-point gain comes from seeing more stain variants or from the feature-level constraints.","The averaged-anchor contrastive idea may transfer to other microscopy domains where appearance factors such as fixation, magnification, or fluorophore can be separated from the structure being classified.","The paper's evaluation does not isolate whether the SNMF decomposition truly preserves morphology, so a direct measurement of morphological consistency under re-staining would be a natural follow-up test."],"forward_implications":["A classifier trained on multi-source blood-cell images can generalize to a new hospital's staining protocol, making automated hematology models potentially deployable without per-site retraining.","Minority classes receive more effective training exposure because each image generates multiple stain variants, which the paper argues helps with rare cell types without harming majority-class accuracy.","The gains are achieved without changing the inference-time architecture or adding extra experts, so the improvement does not come with added deployment cost.","The ablation results show that replacing the averaged-anchor contrastive loss with standard supervised contrastive learning drops F1-macro by 6.4 points, indicating that the domain-invariant anchor construction is a genuine contributor.","The two-stage decoupling of feature learning from classifier training means the method can be bolted onto existing backbones and classification heads."],"supporting_citations":[{"why":"Supplies the sparse non-negative matrix factorization method that separates stain color matrices from density maps in the Beer-Lambert framework.","marker":"[33]"},{"why":"Supplies the decoupling of representation learning and classifier training that SADA adopts to handle data imbalance.","marker":"[10]"},{"why":"Supplies the supervised contrastive learning formulation that SADA modifies into a domain-invariant averaged-anchor loss.","marker":"[29]"},{"why":"BoDA is the main state-of-the-art baseline for imbalanced domain generalization in hematological cell classification.","marker":"[4]"},{"why":"Mixup is a baseline augmentation method for domain generalization that SADA compares against.","marker":"[12]"},{"why":"SelfReg is a self-supervised contrastive regularization baseline that SADA compares against.","marker":"[15]"},{"why":"CORAL is a correlation-alignment baseline that SADA compares against on both public and private datasets.","marker":"[35]"},{"why":"SWAD supplies the weight-averaging model-selection strategy used after training.","marker":"[40]"}],"fun_headline_variants":["Re-staining cells to beat domain shift in blood cell AI","Stain-aware alignment improves blood cell classification across labs","Blood cell AI learns shape, not stain, to transfer better","Synthetic stain transfer tackles domain shift and data imbalance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method relies on the Beer-Lambert sparse factorization cleanly separating stain color from cell structure, so that recoloring the density map with another cluster's stain colors preserves the morphology that defines the cell type.","fun_headline_variants_meta":{"raw":{"variants":["Re-staining cells to beat domain shift in blood cell AI","Stain-aware alignment improves blood cell classification across labs","Blood cell AI learns shape, not stain, to transfer better","Synthetic stain transfer tackles domain shift and data imbalance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000507,"raw_usage":{"total_tokens":2481,"prompt_tokens":962,"completion_tokens":1519,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":1452}},"tokens_in":578,"tokens_out":1519,"duration_ms":14299,"temperature":1.0,"reasoning_tokens":1452,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:53:41.496877+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a single blood-cell image, decompose it via the paper's SNMF procedure, re-stain its density map with stain matrices from the other k-means clusters, and compare a cell-morphology measurement such as nucleus segmentation overlap or cell boundary IoU between the original and re-stained versions; if morphology shifts systematically with the chosen stain matrix, the separation premise fails.","supporting_citations":[{"cited_title":"Structure- preserving color normalization and sparse stain separation for histolog- ical images,","cited_arxiv_id":null,"evidence_quote":"Supplies the sparse non-negative matrix factorization method that separates stain color matrices from density maps in the Beer-Lambert framework."},{"cited_title":"Decoupling representation and classifier for long-tailed recognition,","cited_arxiv_id":null,"evidence_quote":"Supplies the decoupling of representation learning and classifier training that SADA adopts to handle data imbalance."},{"cited_title":"Supervised contrastive learn- ing,","cited_arxiv_id":null,"evidence_quote":"Supplies the supervised contrastive learning formulation that SADA modifies into a domain-invariant averaged-anchor loss."},{"cited_title":"Imbalanced Domain Generalization for Robust Single Cell Classification in Hematological Cytomorphology","cited_arxiv_id":"2303.07771","evidence_quote":"BoDA is the main state-of-the-art baseline for imbalanced domain generalization in hematological cell classification."},{"cited_title":"Adversarial domain adaptation with domain mixup,","cited_arxiv_id":null,"evidence_quote":"Mixup is a baseline augmentation method for domain generalization that SADA compares against."},{"cited_title":"Selfreg: Self-supervised contrastive regularization for domain generalization,","cited_arxiv_id":null,"evidence_quote":"SelfReg is a self-supervised contrastive regularization baseline that SADA compares against."},{"cited_title":"Deep coral: Correlation alignment for deep domain adaptation,","cited_arxiv_id":null,"evidence_quote":"CORAL is a correlation-alignment baseline that SADA compares against on both public and private datasets."},{"cited_title":"Swad: Domain generalization by seeking flat minima,","cited_arxiv_id":null,"evidence_quote":"SWAD supplies the weight-averaging model-selection strategy used after training."}],"review_version":1}