{"id":"fc5fc94e-ee26-454e-819e-a9718555c1aa","arxiv_id":"2411.12350","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DiM adds an f-divergence alignment loss between labeled and unlabeled logits to SAM-based semi-supervised training, claiming state-of-the-art results on two medical segmentation benchmarks.","lead":"This paper combines f-divergence minimization with sharpness-aware optimization for semi-supervised medical image segmentation, reporting improved Dice and boundary metrics on fundus and prostate datasets. A smart generalist might read it to see whether a modest loss-function tweak can push semi-supervised segmentation closer to fully supervised accuracy.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 3.1's Lalign is not the f-divergence defined in Eq. 1 and is not computable from the described Monte Carlo recipe: the density ratio is never estimated, and with 4+4 high-dimensional logits the empirical supports are almost surely disjoint.","rationale":"The reader's weakest assumption is exactly the one I would stress: the f-divergence term is underspecified. I sharpen it: Eq. (2) is not equal to Eq. (1) when supports differ, and a Monte Carlo plug-in over 4+4 empirical logits is degenerate. This does not prove the method fails; the experiments may still be valid if an unstated estimator is used. But the central claim about f-divergence minimization is currently unverifiable. The conditional verdict stands; the condition should be: provide the exact estimator, code, and a zero-ablation of Lalign. The paper does include an upper bound and multiple benchmarks, which is useful, but those do not fill the mechanistic gap.","tokens_in":16104,"tokens_out":8085,"duration_ms":80913,"concrete_test":"Obtain the authors' code and check which objective is actually optimized for Lalign. Then run a minimal diagnostic: on a real 4+4 Fundus minibatch, compute Eq. (2) with the empirical distributions implied by the Monte Carlo description for JS and forward KL, and measure the gradient norm w.r.t. the segmentation head. If the gradient is ~0 (JS, disjoint supports) or NaN (forward KL), the published loss cannot be the training signal; any observed effect must come from an unstated estimator, which the authors must disclose. Alternatively, rerun the Domain 1 Fundus experiment with Lalign set to zero; if Dice/HD/ASD in Table 4 do not change, the f-divergence term is not responsible for the claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism of DiM is the alignment loss Lalign = D_f(plabel∥punlabel) in Eq. (2). Two problems make this term unsupported as the driver of the reported gains. First, Eq. (2) silently drops the f'(∞)plabel(punlabel=0) term that Eq. (1) explicitly adds for non-overlapping supports. For JS divergence f'(∞) = (1/2)log 2 and for Pearson divergence f'(∞) = 1, so whenever the labeled and unlabeled logit supports differ — the regime the paper claims to handle — Lalign is not D_f. Second, no estimator of the density ratio plabel(x)/punlabel(x) is given. The text only says the divergence is 'computed via Monte Carlo estimation based on samples from punlabel.' If plabel and punlabel are empirical distributions over the 4 labeled and 4 unlabeled logit tensors in a batch, the supports are almost surely disjoint in logit space; a plug-in Lalign is then a positive constant for JS (zero gradient) and infinite for forward KL. The introduction also asserts f-divergence 'can still be effectively computed even when the support sets... are different,' but no such estimator is described. Consequently, the f-Divergence rows in Table 4 cannot be attributed to f-divergence minimization without knowing the actual implementation; the gains may arise from an unspecified density-ratio estimator, from SAEM/pseudo-labeling, or from hyperparameter choices. This is a reproducibility and attribution gap at the exact center of the paper's claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DiM, a method for semi-supervised medical image segmentation under domain shift. DiM combines sharpness-aware minimization (SAM) with selective entropy minimization and an alignment loss Lalign that is claimed to minimize the f-divergence between the logit distributions of labeled and unlabeled data. Experiments on Fundus (optic cup/disc) and Prostate datasets report improved Dice/Jaccard and lower surface-distance metrics relative to MiDSS and several other semi-supervised and domain-adaptation baselines, with ablations over SAM and different f-divergence variants. The core mechanism of the paper is the f-divergence alignment term, but its definition and implementation are not fully specified, which affects the reproducibility and attribution of the empirical gains.","tokens_in":16484,"tokens_out":8307,"duration_ms":81719,"significance":"The problem addressed is important: semi-supervised medical image segmentation with data coming from multiple clinical sites is a realistic setting, and improving robustness to domain shift is valuable. If the f-divergence alignment term were properly defined and reliably estimated, the paper would offer a useful extension to SAM-based semi-supervised segmentation. The paper also provides a broad comparison across four fundus domains and six prostate sites, with ablations and visualizations. The main weakness is that the central f-divergence term is not computed as written, no density-ratio estimator is described, and the reported improvements lack variance estimates. These issues are load-bearing for the paper's central claim, so the contribution is currently conditional on substantial revision.","major_comments":[{"comment":"Equation (2) defines Lalign as E_x~p_unlabel[f(p_label(x)/p_unlabel(x))], silently omitting the boundary term f'(∞)p_label(p_unlabel=0) that is explicitly present in Eq. (1). For the Jensen-Shannon and Pearson divergences in Table 1, f'(∞) is nonzero, so whenever the supports of the labeled and unlabeled logit distributions differ — exactly the regime the paper says it handles — Lalign is not equal to D_f. The authors should either retain the boundary term in Lalign or justify why it can be ignored; as written, the central identity Lalign = D_f is false in the intended setting.","section":"Section 3.1, Eq. (1)-(2)"},{"comment":"The paper states that the f-divergence is 'computed via Monte Carlo estimation based on samples from p_unlabel' but never specifies how the density ratio p_label(x)/p_unlabel(x) is estimated. With a batch of 4 labeled and 4 unlabeled samples and high-dimensional logits, the empirical supports of the two distributions are almost surely disjoint; a naive plug-in estimate makes Lalign constant for JS divergence and infinite for forward KL divergence, with zero or undefined gradients. A concrete estimator (e.g., a variational critic, a kernel-based ratio estimate, or a low-dimensional projection with explicit density estimation) is required for the method to be reproducible and for the f-divergence rows in Table 4 to be attributable to f-divergence minimization.","section":"Section 3.1, Monte Carlo estimation"},{"comment":"Tables 2 and 3 report all segmentation metrics as single numbers with no standard deviations, numbers of seeds, or significance tests. The reported gains over MiDSS are modest in places (Fundus average DC 87.28 vs. 85.85; Prostate average DC 87.45 vs. 85.37) and are not consistent across all domains, so the claimed state-of-the-art performance is not statistically supported. Please report mean and standard deviation over at least three independent runs and state the model-selection and seed protocol.","section":"Section 4.3, Tables 2-3"},{"comment":"The 'Upper bound' rows in Tables 2 and 3 are said to be taken from the MiDSS paper, which used UCP within the FixMatch framework with all training data labeled, rather than being computed with the authors' own pipeline, U-Net backbone, SAM optimizer, and data splits. Such an external result cannot serve as a directly comparable fully supervised upper bound for DiM. The authors should either compute their own fully supervised upper bound under identical settings or clearly label the row as an external result and avoid comparing against it as an upper bound for their method.","section":"Section 4.2, upper-bound comparison"}],"minor_comments":[{"comment":"The term Lsym in Eq. (10) is never defined, and the symbol λ appears twice in the same expression; please define Lsym and clarify whether both occurrences denote the same time-dependent weight λ(t).","section":"Section 3.3, Eq. (10)"},{"comment":"The 'baseline' row in Table 4 reports DC 88.27 on Domain 1, whereas the U-Net result with 20 labeled samples in Table 2 is 59.54/73.89; please clarify what the ablation baseline is and how it relates to the main experimental setting.","section":"Section 4.4, Table 4"},{"comment":"UA-MT, FixMatch, CPS, CoraNet, SS-Net, BCP, CauSSL, and MiDSS are semi-supervised methods, not 'supervised techniques' as written in the method list; please correct this description.","section":"Section 4.2, comparison methods"},{"comment":"The sentence beginning 'Directly using gradient norms to filter out unreliable test samples is challenging...' appears twice in consecutive paragraphs; please remove the duplicate.","section":"Section 3.2"},{"comment":"Please proofread for typos and formatting artifacts, including 'utilizef-divergence' at the start of Section 3.1, 'Sencondly' in Section 4.4, and the figure cross-references in Section 4.5, where the text refers to 'Figure 5' for Fundus curves that appear in Figure 4.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The main obstacle to publication is the unspecified f-divergence estimator: without a precise formulation, the reported gains in Table 4 cannot be attributed to f-divergence minimization. I would encourage the editor to require code or a detailed implementation description if the paper is resubmitted. The novelty claim 'first to apply f-divergence constraints to SAM paradigm' should also be checked by the editor, though I have no concrete evidence of prior work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nThe paper you asked about has a solid empirical wrapper but a load-bearing hole at the center. The method combines f-divergence alignment between labeled and unlabeled logits with SAM-based sharpness-aware entropy minimization for mixed-domain semi-supervised medical segmentation. That combination is new in a narrow sense, and the reported numbers on Fundus and Prostate are competitive, often within a point or two of the fully supervised upper bound. The Prostate results in particular are strong.\n\nThe hole is in Section 3.1. Eq. (1) correctly defines f-divergence with the f'(∞) term for non-overlapping supports. Eq. (2) then defines the alignment loss as the same D_f but silently drops that term. If the labeled and unlabeled logit distributions genuinely differ—which is the paper's motivation—the dropped term is nonzero, so L_align is not the divergence they claim. And the paper never says how p_label/p_unlabel is estimated. It says only that the divergence is computed via Monte Carlo from samples. With batch sizes of 4 labeled and 4 unlabeled images, even if you use per-pixel logits, you need a density-ratio estimator or a kernel approximation. None is described. So the f-divergence rows in Table 4 cannot be attributed to f-divergence minimization as defined. The gains could come from the SAM component, the thresholded entropy pseudo-labels, or simply the extra loss's gradient behavior. The stress-test you passed along has it right.\n\nOther soft spots: no error bars, single runs throughout; the upper bound is taken from MiDSS rather than trained by the authors; and the ablation shows the f-divergence alone adds only 0.05 Dice (Table 4 #1 vs #3), which is not convincing. The novelty claim is also overstated: the paper is a straightforward extension of MiDSS with a different regularizer, not a first-principles derivation.\n\nThat said, the paper is not circular, the empirical setup is standard, and the writing is clear enough. The central idea—aligning distributions while seeking flat minima—is reasonable, and the results suggest something works. But as written, the mechanism is not identified, and the formal definition is wrong.\n\nWho gets value from this paper: researchers working on mixed-domain semi-supervised segmentation who might try the combination themselves. It deserves a serious referee, but the referee should demand a precise description of the f-divergence estimator, error bars, and an honest upper bound. I'd suggest sending it to review with the expectation of major revision.\n\n— [Your name]","headline":"The paper's central f-divergence alignment loss, as written, is not the f-divergence defined in Eq. 1 and is not computable without a density-ratio estimator that is never described.","tokens_in":16949,"tokens_out":6065,"would_cite":false,"duration_ms":53499,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DiM adds f-divergence alignment to sharpness-aware optimization and reports higher Dice, Jaccard, and boundary metrics on semi-supervised Fundus and Prostate segmentation than prior methods.","keywords":["f-divergence","sharpness-aware minimization","semi-supervised learning","medical image segmentation","domain shift","feature alignment","entropy minimization","mixed-domain segmentation"],"falsifier":"Run DiM and a SAM-only baseline on the same labeled/unlabeled split while logging the Monte Carlo f-divergence estimate on held-out unlabeled data; if the estimate is high-variance across seeds or stops decreasing while Dice is still improving, or if lowering $L_{\\mathrm{align}}$ does not improve target-domain Dice relative to the SAM-only baseline, the central claim fails. A cleaner control is to replace $L_{\\mathrm{align}}$ with the same loss computed on randomly permuted unlabeled logits: if target-domain performance does not drop, the alignment term is not doing the claimed work.","tokens_in":15942,"feed_emoji":"🩺","tokens_out":6252,"duration_ms":56045,"temperature":0.7,"pith_summary":"The paper proposes DiM, a training method for semi-supervised medical image segmentation that combines sharpness-aware minimization (SAM) with an f-divergence alignment term. The central claim is that minimizing the f-divergence between the logit distributions of labeled and unlabeled images keeps the unlabeled feature distribution close to the labeled one, counteracting feature drift when training and test data come from different domains. On Fundus and Prostate benchmarks with very few labeled samples, DiM reports higher Dice and Jaccard scores and lower boundary distance metrics than previous semi-supervised and unsupervised domain adaptation methods, approaching an upper bound trained on all labels. The intended payoff is a model that performs evenly across source and target domains instead of overfitting to the labeled domain.","feed_headline":"Aligning labeled and unlabeled logits lifts segmentation accuracy","feed_subtitle":"DiM adds an f-divergence term to sharpness-aware training and beats prior semi-supervised methods on Fundus and Prostate.","key_machinery":"The load-bearing object is the f-divergence alignment loss $L_{\\mathrm{align}} = D_f(p_{\\mathrm{label}}\\|p_{\\mathrm{unlabel}}) = \\mathbb{E}_{x\\sim p_{\\mathrm{unlabel}}}[f(p_{\\mathrm{label}}(x)/p_{\\mathrm{unlabel}}(x))]$, evaluated for Jeffrey, Jensen-Shannon, or Pearson divergences via Monte Carlo samples from the unlabeled logits. This term is added to a SAM-based objective that also uses selective entropy minimization, $S(x)E(x;\\theta)$ with an entropy threshold, and a perturbation step $\\hat{\\epsilon}(\\theta) = \\rho\\,\\mathrm{sign}(\\nabla_\\theta E)|\\nabla_\\theta E|/\\|\\nabla_\\theta E\\|_2$ to seek flat minima. The f-divergence term is what carries the cross-domain alignment claim: it penalizes mismatch between the labeled and unlabeled prediction distributions, while the SAM term stabilizes the model under that penalty.","core_discovery":"The paper's discovery claim is that f-divergence minimization is the missing ingredient in sharpness-aware semi-supervised segmentation: SAM flattens the loss landscape, and f-divergence aligns the labeled and unlabeled logit distributions, so the model converges to a flat minimum that is also domain-aligned. The total objective is $L_{\\mathrm{total}} = L_s + \\lambda(L_{\\mathrm{in}} + L_{\\mathrm{out}} + \\lambda L_{\\mathrm{sym}}) + L_{\\mathrm{align}}$, with $L_{\\mathrm{align}} = D_f(p_{\\mathrm{label}}\\|p_{\\mathrm{unlabel}})$ computed by Monte Carlo estimation from minibatches. In the reported experiments, DiM achieves an average Dice of 87.28 on the Fundus optic cup/disc task with 20 labeled samples, compared with 85.85 for MiDSS, and 87.45 on the Prostate task with 40 labeled samples, compared with 85.37 for MiDSS, while also reducing Hausdorff distance and average surface distance. The ablation tables show the best results when SAM and f-divergence are used together, and the t-SNE visualizations show tighter cross-domain clusters than MiDSS.","pith_inferences":["One implicit consequence is that the same alignment loss could be applied to other semi-supervised dense prediction tasks, such as natural-image semantic segmentation or MRI/CT multi-organ segmentation, wherever labeled and unlabeled data come from different acquisition sites.","The paper never specifies how the density ratio $p_{\\mathrm{label}}/p_{\\mathrm{unlabel}}$ is formed from a minibatch of four labeled and four unlabeled logits; a natural extension is to replace the Monte Carlo ratio with a learned density-ratio estimator or a dual f-divergence estimator and test whether small-batch training becomes more stable.","If the benefit of $L_{\\mathrm{align}}$ comes mostly from regularizing the labeled logits rather than from true distribution alignment, then a simpler feature-consistency or entropy-regularization baseline might reproduce the gains; comparing DiM against such baselines with the divergence term ablated would settle that.","The reported convergence curves show stable validation Dice across four fundus domains, which suggests the method may also serve as a diagnostic probe: monitoring the f-divergence estimate during training could indicate when domain shift is being corrected."],"forward_implications":["If DiM's claim is correct, semi-supervised medical image segmentation can be made robust to domain shift without any extra annotations, by adding a single divergence term to an existing SAM-based pipeline.","The reported gains on Fundus (Dice 85.85 to 87.28 with 20 labels) and Prostate (85.37 to 87.45 with 40 labels) suggest that f-divergence alignment recovers a large part of the gap toward fully supervised upper bounds.","Because the method matches or beats dedicated unsupervised domain adaptation baselines without using target labels, it offers a simpler alternative to UDA for mixed-domain clinical data.","The choice of divergence function matters: Jensen-Shannon and Jeffrey variants appear more reliable than Pearson, which gives practitioners a concrete selection rule."],"supporting_citations":[{"why":"Supplies the sharpness-aware minimization method that DiM extends with f-divergence alignment.","marker":"[8]"},{"why":"Defines the mixed-domain semi-supervised segmentation setting and provides the MiDSS baseline and fully-labeled upper bound that DiM compares against.","marker":"[21]"},{"why":"Provides the FixMatch consistency and confidence-threshold framework used for pseudo-label supervision.","marker":"[31]"},{"why":"BCP is a state-of-the-art semi-supervised baseline whose numbers DiM must beat on both benchmarks.","marker":"[2]"},{"why":"CauSSL is a causality-inspired semi-supervised baseline compared in the main tables.","marker":"[23]"},{"why":"SS-Net is a semi-supervised segmentation baseline used in the comparisons.","marker":"[40]"},{"why":"SIFA is an unsupervised bidirectional cross-modality adaptation baseline used for comparison.","marker":"[4]"},{"why":"UDA-VAE++ is an unsupervised domain adaptation baseline for cardiac segmentation used in the comparison.","marker":"[17]"},{"why":"FDA is a Fourier domain adaptation baseline used in the comparison.","marker":"[41]"},{"why":"Provides the KL/Renyi divergence background that motivates choosing f-divergence to avoid infinite values near zero probability.","marker":"[32]"}],"fun_headline_variants":["f-divergence plus sharpness-aware training aligns unlabeled logits","Distribution-aware sharpness improves semi-supervised medical segmentation","Combining SAM and f-divergence yields robust semi-supervised segmentation","f-divergence alignment sharpens the loss landscape for segmentation","Aligning f-divergence with sharpness-aware training for segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the f-divergence between labeled and unlabeled logit distributions can be reliably estimated by Monte Carlo from minibatches of four labeled and four unlabeled samples in high-dimensional logit space, and that minimizing this proxy meaningfully aligns the underlying features.","fun_headline_variants_meta":{"raw":{"variants":["f-divergence plus sharpness-aware training aligns unlabeled logits","Distribution-aware sharpness improves semi-supervised medical segmentation","Combining SAM and f-divergence yields robust semi-supervised segmentation","f-divergence alignment sharpens the loss landscape for segmentation","Aligning f-divergence with sharpness-aware training for segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001788,"raw_usage":{"total_tokens":7093,"prompt_tokens":1036,"completion_tokens":6057,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":652,"completion_tokens_details":{"reasoning_tokens":5968}},"tokens_in":652,"tokens_out":6057,"duration_ms":39951,"temperature":1.0,"reasoning_tokens":5968,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:36:38.239731+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DiM and a SAM-only baseline on the same labeled/unlabeled split while logging the Monte Carlo f-divergence estimate on held-out unlabeled data; if the estimate is high-variance across seeds or stops decreasing while Dice is still improving, or if lowering $L_{\\mathrm{align}}$ does not improve target-domain Dice relative to the SAM-only baseline, the central claim fails. A cleaner control is to replace $L_{\\mathrm{align}}$ with the same loss computed on randomly permuted unlabeled logits: if target-domain performance does not drop, the alignment term is not doing the claimed work.","supporting_citations":[{"cited_title":"Constructing and exploring intermediate domains in mixed domain semi-supervised medical image segmenta- tion","cited_arxiv_id":null,"evidence_quote":"Defines the mixed-domain semi-supervised segmentation setting and provides the MiDSS baseline and fully-labeled upper bound that DiM compares against."},{"cited_title":"Fixmatch: Simplifying semi-supervised learning with consistency and confidence","cited_arxiv_id":null,"evidence_quote":"Provides the FixMatch consistency and confidence-threshold framework used for pseudo-label supervision."},{"cited_title":"Bidirectional copy-paste for semi-supervised medical image segmentation","cited_arxiv_id":null,"evidence_quote":"BCP is a state-of-the-art semi-supervised baseline whose numbers DiM must beat on both benchmarks."},{"cited_title":"Caussl: Causality-inspired semi-supervised learning for medical image segmentation","cited_arxiv_id":null,"evidence_quote":"CauSSL is a causality-inspired semi-supervised baseline compared in the main tables."},{"cited_title":"Exploring smoothness and class-separation for semi-supervised medical image segmentation","cited_arxiv_id":null,"evidence_quote":"SS-Net is a semi-supervised segmentation baseline used in the comparisons."},{"cited_title":"Unsupervised bidirectional cross-modality adaptation via deeply synergistic image and feature alignment for med- ical image segmentation","cited_arxiv_id":null,"evidence_quote":"SIFA is an unsupervised bidirectional cross-modality adaptation baseline used for comparison."},{"cited_title":"Unsuper- vised domain adaptation for cardiac segmentation: Towards structure mutual information maximization","cited_arxiv_id":null,"evidence_quote":"UDA-VAE++ is an unsupervised domain adaptation baseline for cardiac segmentation used in the comparison."},{"cited_title":"Fda: Fourier domain adaptation for semantic segmentation","cited_arxiv_id":null,"evidence_quote":"FDA is a Fourier domain adaptation baseline used in the comparison."},{"cited_title":"R ´enyi divergence and kullback-leibler divergence","cited_arxiv_id":null,"evidence_quote":"Provides the KL/Renyi divergence background that motivates choosing f-divergence to avoid infinite values near zero probability."}],"review_version":1}