{"id":"f7605b30-2172-41c2-aae4-06df5680cfed","arxiv_id":"2411.09434","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A joint latent diffusion model and classifier, sharing one UNet, improves semi-supervised chest X-ray and skin lesion classification and produces counterfactual explanations evaluated by an external classifier and seven doctors.","lead":"Mediffusion trains one diffusion model and one classifier in a shared network, using unlabeled medical images to help classify labeled ones while also generating counterfactual examples that explain the classifier's decisions. A reader should look at this paper if they want a semi-supervised medical classifier with built-in visual explanations rather than a separate explanation tool.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Counterfactual explanations lack a control: CheXNet confidence drops and bounding-box concentration may stem from generic perturbations, not disease-specific features.","rationale":"The reader's conditional verdict already flags several issues, but the most load-bearing is the validity of the counterfactual explanation mechanism, because it is the stated differentiator of the method. The classification results could be acceptable even if explanations fail, but the title and abstract promise 'self-explainable' behavior. My own reading confirms the reader's weakest assumption: nothing in Sec 4.2 rules out that the guidance gradient is exploiting non-clinical shortcuts. The paper does provide positive evidence (CheXNet drop, bbox concentration, doctor preference) but these are confounded. A control experiment is the cleanest way to settle it: if a label-permuted classifier produces the same effect, the explanation is not about disease features at all. I agree with the reader's identification; my concrete test strengthens it by specifying a control. I do not change the verdict: the paper is conditionally acceptable pending this test, and the other secondary correctness issues (ACPL table inconsistency, missing error bars, Eq. 1 mismatch) remain. Thus UNCHANGED.","tokens_in":17884,"tokens_out":6424,"duration_ms":65452,"concrete_test":"Implement the counterfactual removal procedure from Sec 3.5/Appendix F with two control conditions: (A) unguided denoising (guidance scale 0) from t=300; (B) guidance from an auxiliary classifier identical in architecture/training to gω but trained on randomly permuted disease labels. For each condition and for Mediffusion, compute (i) CheXNet target-class confidence drop, (ii) inside-vs-outside bbox mean pixel difference, and (iii) LPIPS distance between original and counterfactual. If (A) achieves ≥30% of Mediffusion's confidence drop, or (B) matches Mediffusion's drop within one standard error, the explanation effect is not attributable to disease-specific features.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Mediffusion's novel 'self-explainable' contribution rests on the claim that classifier guidance ∇_zt log gω(y|eν(zt)) edits disease-relevant features while preserving the rest of the image. The evaluation in Sec 4.2 does not establish this. Tables 4/6 only show that an external CheXNet classifier's confidence for the target class decreases after guided denoising; any perturbation that moves the image away from the target class distribution—brightness, contrast, scanner artifacts—would produce the same effect. The bounding-box concentration (Fig 4) is reported without any baseline; noised-and-denoised images from t=300 could have larger changes inside disease boxes simply because those regions have higher variance in the data, independent of guidance. The doctor survey (Fig 6, Appendix H) rates visibility of disease-indicator changes, not whether the counterfactual is a realistic, minimal, or clinically correct edit. Consequently, the central claim of 'reliable and precise explanations' is not yet supported by the evidence; the observed effects might arise from generic adversarial-like perturbations of the latent code.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Mediffusion, a latent diffusion model whose UNet encoder features are shared with a classifier, trained jointly with a diffusion loss on all data and a classification loss on labeled data. The authors claim that this shared parametrization improves semi-supervised medical image classification and enables self-explainable decisions via classifier-guided counterfactual image edits, in addition to class-conditional synthetic generation. Experiments on ChestXRay14 and ISIC2019 report AUC gains over several SSL baselines, and a counterfactual evaluation uses an externally trained CheXNet and a seven-doctor survey to argue that the edits remove or add disease indicators while preserving other anatomy.","tokens_in":17989,"tokens_out":6797,"duration_ms":64885,"significance":"If the central claims held, the method would be a practically valuable single model for low-label medical imaging: it would combine semi-supervised classification, controllable counterfactual explanations, and guided generation without an external explanation model. The evaluation design has genuine strengths: the counterfactual metrics rely on an independently trained CheXNet classifier and on human expert ratings rather than on the model's own predictions, and the generative assessment uses FID/KID against the target class distribution. However, the current evidence for the explanations' disease specificity is undermined by the absence of control conditions, and the classification comparison contains an internal inconsistency in the main baseline table. The core idea is worth pursuing, but the claims require stronger support.","major_comments":[{"comment":"The ACPL baseline numbers are internally inconsistent. Table 1 reports ACPL without ImageNet pretraining as 63.22/66.68/71.79/74.93 at 2/5/10/20% labels, but Appendix Table 8 lists ACPL*† (no ImageNet) as 72.51/77.10/79.08/80.47 and ACPL*‡ (ImageNet) as 63.22/66.68/71.79/74.93. The main-table row appears to use the ImageNet-pretrained values while labeling them as non-pretrained, or the two tables use different protocol versions. Because this row is the central evidence for the claim of superiority over ACPL without pretraining, the authors must correct the numbers and re-state the comparison.","section":"Section 4.1, Table 1 vs Appendix Table 8"},{"comment":"The joint-model formulation is not consistent with the implemented classifier. Eq. (1) factorizes as pν,ψ,ω(z0:T,y)=pν,ω(y|z0)pν,ψ(z0:T), conditioning classification on the clean latent z0, while Section 3.3 states that gω takes UNet features Ht from the noisy latent zt, and Eq. (2) uses the gradient with respect to zt. The factorization should be written for the actual feature distribution, or the mismatch should be justified. In addition, the final loss L(ν,ψ,ω)=Lclass(ν,ω)−Σ_t L_t,diff(ν,ψ) has a minus sign before a positive diffusion loss; if L_t,diff is the usual MSE noise-prediction loss, minimizing this expression would drive the diffusion loss upward. The sign convention must be corrected.","section":"Section 3.3, Eq. (1)"},{"comment":"The counterfactual evaluation lacks any baseline control, so the central 'self-explainable' claim is not yet supported. The observed drops/increases in CheXNet confidence and the inside-vs-outside bounding-box pixel differences could also be produced by generic perturbations of the latent code, by noise-and-denoise without guidance, or by guidance with a gradient that is unrelated to the disease. The authors should add controls such as (i) unguided noising/diffusion at the same timestep, (ii) guidance toward a random or wrong class, and (iii) guidance from an independent classifier, and report the same metrics (CheXNet confidence change and inside/outside bbox difference) for those controls. Without such controls, Tables 4-5 and Figure 4 do not establish that the edits are disease-specific rather than adversarial-like perturbations.","section":"Section 4.2, Tables 4-5, Fig. 4"},{"comment":"The ablation in Table 3 does not isolate the effect of the diffusion loss. Moving from DenseNet to 'UNet w/o diffusion' changes the architecture, the optimizer settings, the training budget (100 epochs vs. 200,000 steps), and the use of unlabeled data, and then moving to Mediffusion adds the diffusion objective and unlabeled data simultaneously. The claim 'thanks to the additional diffusion loss' requires a controlled comparison in which the architecture, labeled-data protocol, and training budget are fixed and only the diffusion/unlabeled objective is toggled.","section":"Section 4.1, Table 3"},{"comment":"The paper's positioning as comparable to recent semi-supervised methods is weakened by the omission of PEFAT from the main results. Appendix Table 8 shows PEFAT outperforming Mediffusion on ChestXRay14 at every label percentage (e.g., 75.06 vs. 71.85 at 2%), yet Table 1 and the accompanying text do not mention this baseline. Either PEFAT should be included in the main table and the claims reworded, or the authors should explain why they consider their runs directly comparable.","section":"Section 4.1 and Appendix Table 8"}],"minor_comments":[{"comment":"The counterfactual generation procedure is described only in prose; provide a pseudocode algorithm or precise step list with the exact t values and guidance scales in the main text, since these are critical to reproducing the reported behavior.","section":"Section 3.5"},{"comment":"The doctor survey asks about visibility of disease-indicator changes; this measures perceptibility, not correctness, minimality, or clinical validity of the counterfactual. The text should avoid claiming 'practical usability' and 'reliable and precise explanations' solely from this survey.","section":"Section 4.2 and Appendix H"},{"comment":"FID/KID are computed on only 300 generated samples per class with no confidence intervals or seed variance; this is a weak basis for the generative-quality comparison.","section":"Appendix A, Table 11"},{"comment":"The phrase 'guidance scale 100 larger' is ambiguous; specify whether this means adding 100 to the scale or multiplying it.","section":"Appendix H"}],"recommendation":"major_revision","confidential_remarks":"I want to flag for the editor that the main text's comparison table is inconsistent with the appendix and that the counterfactual evaluation would need substantial additional experiments. These issues are fixable but go beyond copy-editing. The paper is not ready for acceptance in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. This is a useful empirical package, not a conceptual breakthrough. The main value is showing that a shared latent-diffusion/classifier UNet can be competitive for semi-supervised ChestXRay and ISIC classification and can produce counterfactual explanations that an external CheXNet and doctors find meaningful. But the headline comparison in Table 1 looks mislabeled, and the counterfactual evaluation lacks a control, so the strongest claims are not yet supported.\n\nWhat is new and what works: the end-to-end setup is concrete, the code is attached, the evaluation is broad, and the doctor survey is a genuine attempt to go beyond automated metrics. Building the classifier into the same parametrization as the denoiser is a natural move, and the attention to low-label regimes is appropriate for the medical setting. The external CheXNet evaluation is a real plus.\n\nThe issues, in proportion. First, the most damaging one: Table 1 reports ACPL at 63.22 for 2% labels, but Appendix Table 8 gives 72.51 for ACPL without ImageNet pretraining and 63.22 for ACPL with pretraining. The main text claims superiority over methods without pretraining, yet the number it prints is the pretrained one. Against the correct no-pretraining column, Mediffusion at 71.85 is below ACPL at 72.51. That error directly affects the central performance claim. Second, PEFAT appears in the appendix and beats Mediffusion on ChestXRay, but it is absent from Table 1. That is not fatal, but it should be in the main comparison. Third, all AUC values are point estimates without error bars, and the Table 3 ablation changes architecture and training budget along with the diffusion loss, so the gain cannot be cleanly attributed to diffusion. Fourth, the counterfactual section needs controls. The stress-test note is right: CheXNet confidence drops and inside-versus-outside bounding-box differences are consistent with target-specific edits, but they are also consistent with generic perturbations. The low “Other diff” values help, and the ACPL comparison in the doctor survey is a decent baseline, but it rates visibility, not minimality or clinical correctness. Add unguided denoising and random-perturbation controls. The Eq. (1) factorization mismatch is real but minor; it should be clarified, not fatal.\n\nThe paper is for people working on semi-supervised medical classification or diffusion-based explainability. It deserves a serious referee. My own verdict is conditional, leaning positive after fixes: correct the table inconsistency, add the missing control conditions, report variance.","headline":"Useful empirical paper with a real table inconsistency that undercuts its main superiority claim; counterfactual evaluation needs controls before the explainability claims can stand.","tokens_in":18627,"tokens_out":3187,"would_cite":true,"duration_ms":35511,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Mediffusion claims that a single latent diffusion model with a shared parametrization can classify medical images with very few labels and also explain its own decisions by generating counterfactual images.","keywords":["semi-supervised learning","latent diffusion models","medical image classification","counterfactual explanations","classifier guidance","chest X-ray","dermoscopy","joint generative-discriminative learning"],"falsifier":"Compute, per disease, the mean absolute pixel change inside expert-annotated bounding boxes versus outside them over many counterfactual pairs; if the inside-outside gap vanishes or flips for a disease while the external classifier still reports a large confidence drop, the explanation is driven by non-clinical features. A second check is to apply the same counterfactual manipulation to images from a different scanner or population and see whether the guidance still localizes to the expected anatomy.","tokens_in":17605,"feed_emoji":"🩻","tokens_out":8662,"duration_ms":74467,"temperature":0.7,"pith_summary":"Mediffusion claims that a single latent diffusion model, with the classifier built on the same UNet that denoises, can do both semi-supervised classification and self-explanations in medical imaging. The paper argues that adding the diffusion denoising objective to the classifier lets the shared representation learn from labeled and unlabeled data alike, so classification improves when labels are scarce. The same shared parameters then guide the denoiser toward or away from a disease class, producing counterfactual images that show what would need to change to flip the prediction. If correct, one model would cover three clinical needs at once: low-label classification, visual explanations, and synthetic data generation.","feed_headline":"A single diffusion model classifies and explains medical images","feed_subtitle":"Sharing one backbone between generation and classification lifts low-label AUC and adds counterfactual explanations.","key_machinery":"The central object is the joint latent diffusion model, in which a single UNet denoiser also serves as the feature backbone of a classifier: the classifier $g_\\omega$ reads the UNet's intermediate features $H_t$ at each diffusion step and outputs class probabilities. The joint objective combines a cross-entropy classification loss on labeled data with the standard diffusion noise-prediction loss on all data. At sampling time, classifier guidance adjusts the noise prediction as $\\hat{\\epsilon}'(z_t) = \\hat{\\epsilon}_\\theta(z_t) - \\sqrt{1-\\bar{\\alpha}_t}\\nabla_{z_t}\\log g_\\omega(y|e_\\nu(z_t))$, which lets the same model generate class-conditioned samples and counterfactuals by guiding toward or against a target class. The counterfactual algorithm adds noise to a latent code up to an intermediate step $t\\in(50,300)$ and denoises with guidance toward the opposite prediction, preserving the original anatomy while altering the disease-indicating regions.","core_discovery":"On the paper's own terms, the discovery is that a joint latent diffusion model with a shared parametrization can learn representations that simultaneously support discriminative and generative tasks, and that the generative path can explain the discriminative one. The authors train a UNet-based latent diffusion model whose intermediate features feed a classifier, optimize the diffusion objective on all data and the classification objective on labeled data, and then use classifier guidance in the latent space to steer denoising. The resulting model reaches a reported AUC of 71.85 at 2% labels on ChestXRay14 against 64.94 for the same UNet without the diffusion objective, and counterfactuals produced by the model lower an external CheXNet classifier's mean confidence for the target disease from 0.70 to 0.30 for Atelectasis while leaving other classes mostly unchanged. The same procedure can also enforce disease indicators on healthy scans and generate class-conditioned synthetic images.","pith_inferences":["I infer that the same joint parametrization could transfer to other low-label settings where unlabeled images are plentiful, such as pathology slides or retinal scans, provided a good latent autoencoder can be trained on the target distribution.","The mechanism suggests a testable diagnostic: if the gradient $\\nabla_{z_t}\\log g_\\omega$ points at clinically meaningful features, then the inside-bounding-box pixel change should correlate with the severity and location of the disease; checking this across many patients would separate true explanation from style artifacts.","Counterfactual generation could double as a label-quality audit: images whose counterfactual toward 'healthy' barely changes the external classifier's confidence may point to mislabeled or ambiguous training samples.","The authors' requirement to keep the noise step small ($50