{"id":"4bca88f3-bed6-4a27-bc99-befcfab4a8d5","arxiv_id":"2411.11087","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"D-Cube combines selected diffusion-model feature maps with ResNet sub-features and custom losses to improve medical image classification.","lead":"This paper introduces D-Cube, a method that combines internal features of a diffusion model with fine-tuned ResNet features to classify medical images from CT, MRI, and X-ray. The authors report accuracy gains over several baselines, especially on small and imbalanced datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing diffusion-free baseline: Table IV's +5.35 gain from fsub means the SOTA result may come from a fine-tuned ResNet head, not diffusion hyper-features; a fsub-only run is needed.","rationale":"The reader's weakest assumption correctly identifies that the load-bearing role of the diffusion features is not established. My stress-test reading confirms this and sharpens it: Table IV's ablation shows the largest gain comes from fsub, yet no diffusion-free fsub-only baseline exists, so the observed SOTA could be produced by a standard fine-tuned ResNet head. The internal contradiction in Table II on the COVID dataset (D-Cube does not beat DiffMIC) reinforces that the empirical claims are overstated. I do not see a reason to change the conditional verdict: the method is plausible and the diffusion components may still help, but the current experiments do not isolate their contribution. A targeted fsub-only comparison, ideally with multi-seed error bars, would settle whether the diffusion hyper-features are genuinely load-bearing. The reader and I agree on the core concern, so the verdict remains conditional pending that experiment.","tokens_in":14749,"tokens_out":4204,"duration_ms":46609,"concrete_test":"Run a fsub-only baseline on all three datasets (pancreas CT, breast MRI, COVID X-ray): train the step-2 classifier using only the fine-tuned ImageNet ResNet sub-features with the same multi-kernel convolution head and the same LCE + LCR + LCycle losses, with no diffusion features, no KS feature selection, and no diffusion-conditioned cycle loss. Report accuracy and F1 over 5 seeds with standard deviations. If fsub-only accuracy is statistically indistinguishable from D-Cube, or if removing diffusion features does not significantly lower performance, the central claim that diffusion hyper-features drive classification is not supported. As a secondary check, repeat the Table VI comparison replacing the KS-selected layers with randomly chosen layers; if random layer sets match the Best Selection within noise, the Gaussianity selection mechanism is not load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central attribution claim is that D-Cube's diffusion hyper-features, selected by a Gaussianity metric, drive the reported classification gains. Table IV does not support this attribution. Starting from a baseline that already includes diffusion features, adding LGen, Fs, and LCls improves accuracy by only +2.96 (85.30 to 88.26), while adding ResNet sub-features fsub improves accuracy by +5.35 (88.26 to 93.61). No experiment trains the step-2 classifier with fsub alone and without diffusion features, so the observed state-of-the-art accuracy could be attributable to the fine-tuned ImageNet-pretrained ResNet head alone. This is not a stylistic concern: the abstract and introduction explicitly credit diffusion hyper-features and feature selection for the improvements, but the experimental design confounds the diffusion contribution with a standard transfer-learning baseline. Table VI further weakens the selection mechanism: Best versus Worst feature selection differs by at most 0.38 accuracy points on any dataset, so the KS-based Gaussianity criterion is not shown to be load-bearing. Additionally, the headline claim that D-Cube surpasses state-of-the-art across all modalities is internally contradicted by Table II on COVID chest X-ray, where DiffMIC achieves 96.40 accuracy and 97.20 F1 versus D-Cube's 96.28 and 96.87. Thus the most load-bearing concern is attribution: the diffusion-specific components may contribute little or nothing beyond the fine-tuned ResNet sub-features.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes D-Cube, a two-stage pipeline for medical image classification. In step 1, a class-conditional U-ViT diffusion model is trained with a contrastive auxiliary loss in addition to the standard denoising objective. In step 2, the diffusion model is frozen, feature maps from layers selected by a Kolmogorov-Smirnov Gaussianity test are fed into a small convolutional classifier, and training is regularized by a cycle loss and a consistency loss. An ImageNet-pretrained ResNet is fine-tuned to provide sub-features that are concatenated with the diffusion features. The method is evaluated on pancreas CT, breast MRI, and COVID chest X-ray datasets against CNN, transformer, and diffusion-based baselines, with additional ablation studies, synthetic-data augmentation experiments, and FID comparisons.","tokens_in":15201,"tokens_out":6065,"duration_ms":63720,"significance":"If the attribution were established, D-Cube would be a valuable demonstration that diffusion-model hyper-features, selected by a Gaussianity criterion, improve medical classification under small-sample and imbalanced conditions. The paper has notable strengths: it validates across three imaging modalities and populations, it includes a fairly broad baseline table, it releases code, and it provides an explicit feature-selection criterion together with FID evidence that the proposed generative loss improves image fidelity. The central problem is that the experiments do not isolate the diffusion contribution from the fine-tuned ResNet sub-features, and the headline claims are stronger than the data support, particularly on the COVID dataset where the main diffusion-based baseline scores higher. With additional controlled ablations and variance reporting, the contribution could be made convincing.","major_comments":[{"comment":"The ablation does not include a diffusion-free baseline, so the paper's central attribution claim is not supported. The largest single accuracy gain in Table IV comes from adding the ResNet sub-features fsub (+5.35 over +LCls), and every full D-Cube model in Table II uses fsub. A controlled run that trains the same step-2 classifier with fsub only, and without any diffusion features, is necessary to determine whether the state-of-the-art results are driven by the diffusion hyper-features or by a fine-tuned ImageNet-initialized ResNet head. Without this baseline, the abstract and introduction's claim that diffusion features and feature selection are responsible for the improvements is confounded.","section":"Section IV-D, Table IV"},{"comment":"The claim that D-Cube 'surpasses previous models in almost metrics across datasets' is contradicted on the COVID chest X-ray dataset: DiffMIC achieves 96.40 accuracy and 97.20 F1 versus D-Cube's 96.28 and 96.87. The text in Section IV-C also states that 'by utilizing sub-features, we outperformed DiffMIC,' which is not true for this dataset. Please revise the claim, discuss the failure, and report whether the comparison changes under multiple seeds or a paired statistical test.","section":"Section IV-C, Table II"},{"comment":"All reported numbers appear to be from single training runs, with no standard deviations, confidence intervals, or significance tests. This is particularly concerning on the Duke MRI dataset, which has only 707 test images, where the differences between the best and worst feature selections are small (Table VI) and the ablation increments are small (Table IV). I request that the main comparisons and ablations be repeated for at least 3-5 seeds and reported as mean +/- std, with a paired test such as McNemar's test or bootstrapping for the head-to-head comparisons.","section":"Section IV, Tables II, IV, and VI"},{"comment":"The proposed KS-based Gaussianity feature selection is not shown to be load-bearing. Replacing the 'best' layers with the 'worst' layers changes accuracy by only 0.38, 1.14, and 0.43 points on the pancreas, breast, and COVID datasets, respectively. The manuscript also does not specify how many layers were selected, which layer indices correspond to the yellow stars and red crosses in Fig. 2, or how stable the selection is across random batches, timesteps, or model initializations. Please provide this information and compare the KS-selected layers against a random selection of the same number of layers, or a fixed high-level/low-level layer choice, to establish that the criterion is more than a negligible detail.","section":"Section III-B2a and Table VI"},{"comment":"The method description is internally inconsistent about the classifier input. Section IV-B2 says step 2 uses 'the original image x0 without any noise at a random time step t as input,' while the cycle loss in Eq. (7) is defined in terms of xt and the ground-truth and predicted class labels, and Table VII explicitly varies between x0 and xt. Please clarify exactly what input the classifier receives, what role t plays in step 2, and how Eq. (7) is computed when the classifier sees x0 rather than xt.","section":"Section IV-B2 and Eq. (7)"},{"comment":"The abstract and introduction make unqualified claims of 'state-of-the-art diagnostic accuracy and efficiency' and 'superior performance compared to existing baseline models.' These statements should be qualified to the datasets and metrics where the comparisons actually hold, especially in light of the COVID result in Table II and the missing diffusion-free baseline in Table IV.","section":"Abstract and Section I"}],"minor_comments":[{"comment":"The contrastive loss in Eq. (4) is written for a batch but does not explain how pairs (x1_t, x2_t) are sampled: are they two augmented views of the same image, two different images from the same class, or random pairs? Please specify the pairing rule.","section":"Eq. (4)"},{"comment":"The text says 'features from layers marked with yellow star, corresponding to p-values less than 0.05, were considered the best selections,' but the feature-selection section says that p-values greater than 0.05 indicate Gaussian maps. The logic is consistent only if the reader understands that non-Gaussian maps are semantically informative; please state this explicitly and use consistent notation for L_Diff vs 'Ldiff.","section":"Section IV-D1"},{"comment":"The sentence 'D-Cube surpasses previous models in almost metrics across datasets' should read 'almost all metrics,' and the surrounding paragraph should be rewritten to distinguish the datasets on which D-Cube is best from the one where it is second-best.","section":"Section IV-C"},{"comment":"The table is titled 'Augmentation with synthetic data of D-Cube,' but its rows are ResNeXt-101 and ViT-B/16 augmented with synthetic images. Please retitle the table to indicate that it evaluates the effect of synthetic-data augmentation on baseline models, not on D-Cube.","section":"Table III"},{"comment":"The F1 value for ResNeXt-101 sub-features is reported as 89.6, whereas the corresponding full D-Cube entry in Table II reports 89.69. Please correct the inconsistency and report precision and recall with consistent decimal places.","section":"Table VIII"},{"comment":"The caption of Fig. 2 does not define the axes, the exact layers, or how the 'best' and 'worst' combinations were formed. Please label axes, indicate the number of layers selected, and describe in the caption how the yellow-star and red-cross markers map to the entries in Table VI.","section":"Fig. 2 and Table VI"}],"recommendation":"major_revision","confidential_remarks":"The paper is on a timely topic and contains useful components, but the current evidence is insufficient for the strength of the claims. The missing diffusion-free baseline and the absence of any variance estimation are the two issues that most need to be addressed before acceptance. I would also encourage the authors to soften the state-of-the-art claim until the COVID comparison is resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline claim is not supported by the paper's own data. On pancreas CT, the diffusion components (LGen, feature selection, classification losses) add about 2.96 points (85.30 to 88.26), while adding the ResNet sub-features fsub adds 5.35 points (88.26 to 93.61). There is no experiment that trains the classifier with fsub alone and no diffusion features, so the state-of-the-art result could be mostly a standard transfer-learning effect. The abstract and introduction credit diffusion hyper-features, but Table IV tells a different story.\n\nTo be fair, there is real content here. The Gaussianity-based layer selection (KS test on feature maps) is a new heuristic, and the idea of selecting non-Gaussian layers as semantic is worth testing. The contrastive diffusion loss LGen does lower FID, and synthetic data from the trained diffusion model improves weaker baselines (ResNeXt, ViT) by 1–4 points on recall and F1. Evaluating on CT, MRI, and X-ray across Asian and Western populations is a plus. The writing is clear, and the ablation is more informative than most.\n\nThe soft spots are significant. First, the missing fsub-only baseline is the load-bearing gap. Second, Table VI shows Best vs Worst feature selection differs by 0.38 to 1.14 accuracy points depending on the dataset, so the KS criterion is not shown to be load-bearing. Third, the COVID SOTA claim is directly contradicted by Table II: DiffMIC beats D-Cube on all four metrics for COVID (Acc 96.40 vs 96.28, F1 97.20 vs 96.87). Fourth, there are no error bars or multi-seed results, so we cannot tell whether any of these gaps are meaningful. Fifth, the KS test is computed on a single batch of 256 images at timestep t=100, with no sensitivity analysis. Finally, the cycle loss weight varies a lot across datasets (10 vs 100), suggesting the loss is heavily tuned.\n\nWho this is for: people working on diffusion features for medical imaging, and anyone benchmarking transfer-learning baselines. The paper deserves a serious referee, but it needs major revision: add the fsub-only baseline, report variance, correct the COVID claim, and soften the attribution language. As is, I would not cite it as evidence for diffusion-feature superiority.","headline":"The paper's own ablation undercuts its central claim: the largest gain comes from a fine-tuned ResNet sub-feature stream, not from diffusion hyper-features.","tokens_in":15575,"tokens_out":3078,"would_cite":false,"duration_ms":80683,"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":"D-Cube claims that frozen diffusion-model features, selected by testing which layers deviate from Gaussian noise, set new high scores on pancreas CT, breast MRI, and COVID chest X-ray classification.","keywords":["diffusion models","medical image classification","feature selection","Gaussianity metric","contrastive learning","cancer diagnosis","synthetic data generation"],"falsifier":"Run the layer-selection step again on a different random batch or a different timestep and check whether the same layers are flagged as non-Gaussian; if the selected layers change, the metric is not a stable property of the model. Also test a version where the selected diffusion features are replaced by the same number of randomly chosen layers: if random layers match the KS-selected accuracy, the Gaussianity criterion is not doing the work.","tokens_in":14577,"feed_emoji":"🩺","tokens_out":7926,"duration_ms":68046,"temperature":0.7,"pith_summary":"The paper sets out to show that a diffusion model can do more than generate synthetic medical images: its internal feature maps, once the noise-dominated layers are filtered out, carry the semantic information needed for cancer classification. The proposed D-Cube pipeline first trains a class-conditional diffusion model with an added contrastive loss, then freezes it and trains a classifier on a concatenation of selected diffusion features and fine-tuned CNN sub-features. On pancreas CT, breast MRI, and COVID chest X-ray, the paper reports accuracies of 93.61%, 77.98%, and 96.28%, beating the compared CNN, transformer, and diffusion baselines on almost all metrics. The ablation attributes the largest single gain to the CNN sub-features, with the Gaussianity-based layer selection and the tailored losses adding smaller but consistent improvements.","feed_headline":"Diffusion features, chosen by a noise test, top medical benchmarks","feed_subtitle":"Frozen diffusion model plus CNN sub-features hits 93.6% on pancreas CT and 78.0% on breast MRI.","key_machinery":"The load-bearing object is the Gaussianity metric for layer selection: a Kolmogorov-Smirnov test that compares the empirical cumulative distribution of each diffusion layer's feature map, computed over a batch of 256 images at timestep t=100, with the CDF of a Gaussian. Layers with p-value above 0.05 are judged to be doing noise prediction and are excluded; layers with p-value below 0.05 are judged to preserve semantic image content and are kept. The argument is that a denoising network must contain both kinds of layers, and that the non-Gaussian ones are exactly the features worth feeding to a classifier. Around this selection, the pipeline adds a contrastive loss during diffusion training, a cycle loss that ties classifier predictions back to the frozen diffusion model's noise estimates, a consistency-regularization loss on flipped inputs, and a channel-expansion step that multiplies the single-channel diffusion features by fine-tuned CNN sub-features.","core_discovery":"D-Cube's central claim is that intermediate features of a class-conditional denoising diffusion model, trained with a contrastive loss and then frozen, are strong enough to drive medical image classification once the layers that merely predict noise are discarded. The paper identifies those layers with a Kolmogorov-Smirnov test: a feature map whose p-value against a Gaussian distribution is above 0.05 is treated as noise and dropped, while non-Gaussian maps are kept as semantic hyper-features. These selected features are concatenated, multiplied channel-wise by sub-features from a fine-tuned CNN, and classified by a small convolutional head trained with cross-entropy, a cycle loss that checks whether the predicted class reproduces the ground-truth diffusion noise, and a consistency-regularization loss on horizontal flips. The paper reports that this combination scores 93.61% accuracy on pancreas CT, 77.98% on breast MRI, and 96.28% on COVID X-ray, outperforming the compared baselines on almost all metrics and exceeding the diffusion baseline DiffMIC on the two cancer datasets.","pith_inferences":["A natural next test is whether the same KS-based layer selection transfers to other generative backbones, such as latent diffusion models, where the noise-dominated and semantic layers may be arranged differently.","Because the feature-selection p-values are computed on a single batch and a single timestep, a more robust version of the method would average the Gaussianity statistic over several batches and timesteps; if the selected layers shift across those repetitions, the selection criterion is less principled than the paper implies.","The ablation showing that CNN sub-features add 5.35 accuracy points suggests that a fair head-to-head should compare D-Cube against a classifier using only those sub-features with the same loss functions, to see how much of the final score is actually due to the diffusion hyper-features.","The method's claim of broad applicability could be probed on other modalities with similar imbalance, such as ultrasound or histopathology, where the Gaussianity pattern of diffusion layers may differ."],"forward_implications":["If the central claim holds, a frozen diffusion model pretrained on medical data can be reused as a classification feature extractor, avoiding the need to fine-tune the generator for each new task.","The Gaussianity metric gives a training-free rule for choosing layers, replacing hand-picked layer and timestep choices that other diffusion-feature methods rely on.","The reported gains on the smaller breast MRI dataset suggest the method is most valuable when data are scarce and class boundaries are blurred.","Synthetic images produced by the contrastively trained diffusion model can be added to training sets, improving standard CNN and transformer classifiers on the pancreas dataset by up to 2-4% in recall and F1.","The ablation's large jump from the CNN sub-features implies that the diffusion features and the CNN features make complementary rather than redundant contributions."],"supporting_citations":[{"why":"Supplies the DDPM forward/reverse formulation whose intermediate features D-Cube extracts and classifies.","marker":"[3]"},{"why":"U-ViT is the backbone architecture used as the diffusion model in the paper's Step 1.","marker":"[9]"},{"why":"DiffMIC is the main diffusion-based classification baseline that D-Cube compares against on all three datasets.","marker":"[29]"},{"why":"ResNet is the pretrained CNN whose fine-tuned sub-features produce the largest accuracy gain in the ablation.","marker":"[12]"},{"why":"One of the two references for the COVID-19 chest X-ray dataset used in evaluation.","marker":"[30]"},{"why":"The companion COVID-19 chest X-ray dataset reference used alongside [30] for evaluation.","marker":"[31]"},{"why":"Supplies the Duke breast cancer MRI dataset with patient-level train/test split used for evaluation.","marker":"[32]"},{"why":"Prior evidence that diffusion features contain semantic correspondence, motivating the search over layers and timesteps.","marker":"[46]"},{"why":"Shows diffusion-model features can support downstream tasks with few labels, supporting the feature-extraction strategy.","marker":"[47]"},{"why":"Shows synthetic data from diffusion models can improve downstream classification, grounding the augmentation experiments.","marker":"[8]"}],"fun_headline_variants":["KS test prunes diffusion noise, D-Cube hits 93.6% on pancreas CT","D-Cube: non-Gaussian diffusion maps + CNN sub-features top cancer benchmarks","D-Cube: KS-filtered diffusion features beat DiffMIC on CT and MRI","D-Cube: contrastive diffusion plus KS test selects features for 93.6% pancreas","D-Cube: KS-selected diffusion features score 96.3% on COVID X-ray"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method rests on the premise that layers whose feature maps deviate from a Gaussian distribution are the ones holding classification-relevant semantics, and that a single batch of 256 images at timestep t=100 is enough to tell which layers those are.","fun_headline_variants_meta":{"raw":{"variants":["KS test prunes diffusion noise, D-Cube hits 93.6% on pancreas CT","D-Cube: non-Gaussian diffusion maps + CNN sub-features top cancer benchmarks","D-Cube: KS-filtered diffusion features beat DiffMIC on CT and MRI","D-Cube: contrastive diffusion plus KS test selects features for 93.6% pancreas","D-Cube: KS-selected diffusion features score 96.3% on COVID X-ray"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001243,"raw_usage":{"total_tokens":5102,"prompt_tokens":948,"completion_tokens":4154,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":4038}},"tokens_in":564,"tokens_out":4154,"duration_ms":31130,"temperature":1.0,"reasoning_tokens":4038,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:54:38.735688+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the layer-selection step again on a different random batch or a different timestep and check whether the same layers are flagged as non-Gaussian; if the selected layers change, the metric is not a stable property of the model. Also test a version where the selected diffusion features are replaced by the same number of randomly chosen layers: if random layers match the KS-selected accuracy, the Gaussianity criterion is not doing the work.","supporting_citations":[{"cited_title":"All are worth words: A vit backbone for diffusion models,","cited_arxiv_id":null,"evidence_quote":"U-ViT is the backbone architecture used as the diffusion model in the paper's Step 1."},{"cited_title":"I., Schönlieb, C","cited_arxiv_id":null,"evidence_quote":"DiffMIC is the main diffusion-based classification baseline that D-Cube compares against on all three datasets."},{"cited_title":"Can AI help in screening Viral and COVID-19 pneumonia?","cited_arxiv_id":null,"evidence_quote":"One of the two references for the COVID-19 chest X-ray dataset used in evaluation."},{"cited_title":"Exploring the Effect of Image Enhancement Techniques on COVID-19 Detection using Chest X-rays Images","cited_arxiv_id":"2012.02238","evidence_quote":"The companion COVID-19 chest X-ray dataset reference used alongside [30] for evaluation."},{"cited_title":"and Mazurowski, M.A., 2018","cited_arxiv_id":null,"evidence_quote":"Supplies the Duke breast cancer MRI dataset with patient-level train/test split used for evaluation."},{"cited_title":"Diffusion hyperfeatures: Searching through time and space for semantic correspondence","cited_arxiv_id":null,"evidence_quote":"Prior evidence that diffusion features contain semantic correspondence, motivating the search over layers and timesteps."},{"cited_title":"Label-Efficient Semantic Segmentation with Diffusion Models","cited_arxiv_id":null,"evidence_quote":"Shows diffusion-model features can support downstream tasks with few labels, supporting the feature-extraction strategy."}],"review_version":1}