{"id":"310a6f4c-3ef1-438b-8449-14b48d6ca90a","arxiv_id":"2506.15218","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DM-FNet is a two-stage diffusion-based network that fuses multimodal medical images with a single parameter set, reporting top scores on most of nine benchmark metrics.","lead":"DM-FNet trains a UNet with a diffusion process, then uses its multiscale features to fuse medical images from different modalities with one set of network weights. The paper reports strong scores on standard fusion metrics across MRI-CT, MRI-PET, and MRI-SPECT, plus two extra tasks, but key settings were tuned on the test sets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-set tuning of α, β, and time steps in §IV-B inflates the reported SOTA comparison; a held-out validation split is required before the central claim can be assessed.","rationale":"The diffusion construction itself is not internally inconsistent: the forward and reverse formulations in Section III are standard, and the 'w/o Dif.' ablation provides some evidence that the diffusion-trained encoder contributes. The load-bearing weakness is in the evaluation protocol, not in the equations. The strongest claim depends on the experimental comparisons, and those comparisons are not trustworthy because the test set was used for model selection. This is a validity condition, not an accusation of bad faith. The reader's rationale already flags test-set-based hyperparameter selection, which is why my agreement is partial rather than full: the reader's stated weakest assumption concerns the informativeness of the diffusion features, whereas the more decisive vulnerability is that α, β, and the time-step set were selected on the same 50-pair test sets used for the final comparison. The proposed check would settle whether the reported SOTA margins survive an honest model-selection procedure. If they do, the central claim stands; if they shrink, the paper would need to be reframed as a weaker, configuration-dependent result. Since the manuscript is otherwise coherent and the code is promised, the appropriate disposition remains the same conditional verdict: accept only after a proper validation split and variance reporting are supplied.","tokens_in":20465,"tokens_out":4106,"duration_ms":46060,"concrete_test":"Hold out a validation split from the 90 training pairs per modality (e.g., 70 train / 20 validation, with no overlap with the 50-pair test sets). Repeat the Table I and Table II ablations and choose α, β, and time steps from validation only; freeze the selected configuration, retrain, and recompute Tables V and VII on the untouched test sets. Report mean and std over at least three seeds. If the selected hyperparameters change or the test-set margins over the SOTA baselines shrink appreciably, the headline performance claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — a single set of parameters outperforming nine state-of-the-art methods on most metrics — rests on the comparisons in Tables V and VII. These comparisons are compromised because the same 50-pair test sets were used to select the model configuration. Section IV-B reports the hyperparameter analysis 'averaged over three test sets' (Table I) and the time-step analysis 'averaged over three test sets' (Table II), and those tables are the basis for fixing α=1.5, β=0.5 and t1=5, t2=10, t3=20 (Section IV-A). The final metrics in Tables V/VII and Figs. 7/10 are then computed on exactly the same test sets. This makes the test set a validation set: the reported margins over the nine baselines may reflect selection on the evaluation target rather than a genuine advantage. No error bars or repeated-seed statistics are provided, and the training set is only 90 pairs, so the comparison is also noisy. A reader cannot currently tell whether the diffusion-trained multi-time-step encoder is the cause of the improvement or whether the reported configuration was chosen because it scored well on the test data.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DM-FNet, a two-stage framework for multimodal medical image fusion. Stage I trains a UNet denoiser with a diffusion process for image reconstruction; Stage II feeds the trained UNet with source images corrupted by Gaussian noise at three selected time steps (t=5, 10, 20) and uses attention-based multimodal feature fusion (AMFF) and multiscale feature fusion (MSFF) to combine encoder-decoder features and produce the fused image. A hybrid loss with intensity, SSIM, and gradient terms is used. Experiments on MRI-CT, MRI-PET, and MRI-SPECT datasets compare against nine methods; the authors also report generalization to PC-GFP without fine-tuning and to infrared-visible fusion after fine-tuning. The code is publicly available.","tokens_in":20696,"tokens_out":6583,"duration_ms":57193,"significance":"If the evaluation were clean, the contribution would be meaningful: the paper introduces a nontrivial two-stage design that uses diffusion training as a feature extractor and demonstrates a single parameter set across three medical fusion tasks, with public code and ablations. The multi-time-step feature fusion and the module design are clearly described, and the extension to PC-GFP without fine-tuning is a good generalization test. However, the current evidence for the central claim is weakened by evaluation-protocol issues that must be resolved before the reported improvements can be attributed to the method rather than to selection on the evaluation target.","major_comments":[{"comment":"The hyperparameter analysis in Section IV-B selects the intensity/SSIM weights (α=1.5, β=0.5) and the time-step combination (t1=5, t2=10, t3=20) by averaging metrics over the three test sets (Tables I and II), and then Tables V and VII report final comparisons computed on the same test sets. Because the test sets were used to choose the configuration, the reported margins over the nine baselines may reflect selection on the evaluation target rather than a genuine advantage. A held-out validation split or nested cross-validation is required before the state-of-the-art claim can be assessed.","section":"Section IV-A/B (Tables I and II)"},{"comment":"The ablation labeled \"w/o Dif.\" removes the diffusion process in Stage I and simultaneously removes the multi-time-step noisy inputs in Stage II (\"the fusion module of different noise conditions was removed\"). This conflates the effect of diffusion-based training with the effect of the multi-time-step input strategy. To attribute the improvement to diffusion-process training, the paper should compare against a control in which a conventionally trained UNet is fed the same multi-time-step noisy inputs, or in which the diffusion-trained UNet is used with a single time step.","section":"Section IV-B (Table III)"},{"comment":"The central comparison rests on point estimates over a single train/test split with only 90 training pairs and 50 test pairs per modality. Several winning margins are small (e.g., Table V: Proposed SCD 1.508 vs Cloud 1.378 and VIFF 0.600 vs 0.570; Table VII: Proposed Q_W 0.925 vs SHIP 0.926, where SHIP is slightly higher). No error bars, confidence intervals, or significance tests are reported, so the \"outperforming\" claim is not yet robust. At minimum, the authors should report results over multiple seeds or provide statistical tests.","section":"Section IV-C (Tables V and VII)"},{"comment":"The forward diffusion process is defined with a total of T steps and a variance schedule α_t (Eqs. (1)-(2)), and Stage II uses t=5, 10, 20, but the paper never reports the value of T or the specific schedule used. Because the diffusion-trained UNet is the foundation of the feature extractor, the absence of these values prevents reproduction of Stage I and should be corrected.","section":"Section III-C and IV-A"}],"minor_comments":[{"comment":"The section heading \"Loss fuction\" should read \"Loss function.\"","section":"Section III-E"},{"comment":"The acronym \"SOAT\" appears where \"SOTA\" (state of the art) is intended; please correct it throughout, including \"nine SOAT methods\" in the experimental sections and figure captions.","section":"Section IV-A and throughout"},{"comment":"The comparative method is introduced as \"TL-SR\" in Section IV-A but is referred to as \"TLSR\" in later sections and tables; please unify the naming.","section":"Section IV-A and Section IV-C"},{"comment":"The sentence \"Most of the evaluation metrics across all test sets in Table V were also noticeably reduced\" appears to refer to Table III, not Table V; please correct the cross-reference.","section":"Section IV-B (paragraph on the diffusion process)"},{"comment":"The statement \"removing any module results in a significant performance drop\" is not fully supported by Table IV: for MRI-CT, \"w/o MSFF\" yields higher AG (9.237 vs 9.000) and higher Q_AB/F (0.586 vs 0.571), and \"w/o AMFF\" yields higher Q_AB/F (0.588 vs 0.571). Please qualify this claim or explain how these increases are consistent with the stated conclusion.","section":"Section IV-B (Table IV discussion)"},{"comment":"The notation IY_{B,t} introduces a superscript Y that is not defined when I_B is CT (grayscale). Clarify how the YCbCr luminance channel is used for CT as well as for PET/SPECT inputs.","section":"Equation (4) and color-space descriptions"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the two-stage diffusion-based fusion idea is worth pursuing, but the evaluation protocol is the key weakness. The test-set-based hyperparameter and time-step selection followed by evaluation on the same test sets needs to be redone with a proper validation split, and statistical robustness should be addressed. I recommend requesting a major revision with these points as mandatory changes."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read on arXiv:2506.15218 (DM-FNet). The core idea is worth a look: train a UNet with a diffusion objective, then freeze it and feed the fusion network noisy images at multiple time steps (5, 10, 20) so it sees features at different noise levels. That multi-time-step mix plus a single parameter set across MRI-CT, MRI-PET, and MRI-SPECT is genuinely new relative to Dif-Fusion, which trains a diffusion feature extractor but only for infrared-visible and doesn't do the time-step ensemble. The ablations in Section IV-B are the best part: they show the diffusion training, AMFF, and MSFF all contribute, and the PC-GFP result without fine-tuning suggests the representation generalizes.\n\nNow the soft spot, and it's a real one. The hyperparameters alpha, beta, and the time steps are selected by looking at Table I and Table II, which are averaged over the three 50-pair test sets. The final comparisons in Tables V and VII are on the same test sets. So the test set is doing double duty as a validation set. That means the reported margins over the nine baselines are inflated by selection. The paper would need a proper held-out validation split (or an inner CV loop) to claim SOTA. There are also no error bars or repeated-seed results, and the training set is only 90 pairs, so the comparison is noisy. I'd also like a per-dataset breakdown in Table V instead of pooling MRI-PET and MRI-SPECT into one ranking.\n\nNone of this kills the method. The architecture is plausible, the ablations are informative, and the efficiency story (0.329s per pair, no iterative sampling) is credible. But the headline claim should be read as 'promising, not proven.' The code link is there, though I haven't verified it runs.\n\nWho's this for? Anyone working on medical image fusion or diffusion-based feature extraction. It's a workshop-level or decent journal paper, but not a definitive SOTA result as written. I'd send it to peer review with a request that the authors redo the hyperparameter selection on a validation split and report variance; I would not desk-reject it. For my own work, I wouldn't cite it as a reference result until that's fixed.","headline":"Useful idea with a real methodological flaw: the SOTA claim rests on hyperparameters chosen on the same test sets used for the final comparison, so it needs a validation split before the numbers can be trusted.","tokens_in":21211,"tokens_out":2179,"would_cite":false,"duration_ms":23416,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A two-stage diffusion-trained network fuses MRI-CT, MRI-PET, and MRI-SPECT images with a single set of parameters.","keywords":["medical image fusion","diffusion model","encoder-decoder","attention mechanism","multiscale fusion","unified fusion","image reconstruction","multimodal imaging"],"falsifier":"Retrain the Stage I encoder with the same UNet and a plain reconstruction loss (no diffusion schedule) while keeping Stage II identical, and compare on the three test sets; if the non-diffusion encoder matches or exceeds DM-FNet on the nine metrics, the diffusion process itself is not the load-bearing component.","tokens_in":20251,"feed_emoji":"🧠","tokens_out":5430,"duration_ms":49362,"temperature":0.7,"pith_summary":"This paper claims that a two-stage network, DM-FNet, can fuse structurally and functionally different medical image pairs (MRI-CT, MRI-PET, MRI-SPECT) with one set of parameters. The first stage trains a UNet through a denoising diffusion process, so the encoder learns fine anatomical detail through progressive denoising. The second stage feeds the fusion network the same source images corrupted by Gaussian noise at several time steps, then combines multiscale, cross-modal features with attention-guided fusion modules. On three medical test sets, the paper reports that this design outperforms nine previous fusion methods on most of nine quality metrics, and that the same model transfers to phase-contrast and green-fluorescent-protein fusion without fine-tuning. The appeal is a single unified fusion model rather than task-specific weights.","feed_headline":"Diffusion-trained network beats nine fusion methods on most metrics","feed_subtitle":"One parameter set fuses MRI-CT, MRI-PET, and MRI-SPECT pairs without task-specific retraining.","key_machinery":"The load-bearing mechanism is the diffusion-trained UNet encoder used as a fixed feature extractor. In Stage I, a denoising diffusion probabilistic model trains the UNet to predict the previous denoising step, forcing it to represent multilevel image structure; in Stage II the fusion network receives source images with Gaussian noise added at time steps 5, 10, and 20, so the same encoder sees each image at three noise levels. The attention-guided multimodal feature fusion module (AMFF) computes spatial, channel, and pixel attention maps that weight each modality's features, and the multiscale feature fusion module (MSFF) hierarchically combines the five encoder scales. The hybrid loss, which combines an intensity term, a maximum-local-standard-deviation SSIM term, and a Sobel gradient term, drives the final trade-off between luminance and detail.","core_discovery":"The central claim is that diffusion-process training can serve as a feature-extraction engine for image fusion, not as a slow generative sampler. DM-FNet first trains a UNet to reconstruct clean medical images by removing Gaussian noise step by step; that training forces the network to encode the fine structures that noise masks, such as bone edges, soft tissue boundaries, and tumor margins. The fusion stage then corrupts each source image at time steps 5, 10, and 20, encodes those noisy versions with the frozen diffusion-trained UNet, and merges the resulting multilevel features using three modules: a feature-fusion step that combines the different time-step encodings, an attention-guided multimodal feature fusion module (AMFF) that blends spatial, channel, and pixel attention across the two modalities, and a multiscale feature fusion module (MSFF) that integrates the five UNet scales. A hybrid loss combining intensity, SSIM, and gradient terms balances brightness and detail. The paper reports that this unified architecture achieves the best or second-best scores on most of nine metrics across MRI-CT, MRI-PET, and MRI-SPECT, and that removing the diffusion process or either fusion module measurably degrades results.","pith_inferences":["One testable extension is whether the multi-time-step noise schedule acts purely as data augmentation: a simpler stochastic noise augmentation without the diffusion reconstruction objective might match DM-FNet if the diffusion objective itself is not essential.","Feeding noisy inputs at inference time is unusual; comparing DM-FNet against the same fusion network fed only clean source images would isolate whether the noise corruption, rather than the diffusion-trained weights, drives the reported gain.","Because nothing in the method is medical-specific except the training data, the same design could transfer to other fusion domains such as multi-focus, multi-exposure, or remote-sensing fusion where one unified model is desirable."],"forward_implications":["One trained parameter set suffices for MRI-CT, MRI-PET, and MRI-SPECT fusion, so new clinical fusion tasks need not retrain a dedicated model.","Because the diffusion forward process runs only in training, fusion at test time avoids the slow iterative sampling of generative diffusion fusers such as DDFM or FusionDiff.","The same fusion network extends to infrared-visible fusion after fine-tuning and to PC-GFP fusion without fine-tuning, indicating that the diffusion-trained features generalize across modality types.","Removing the diffusion process or either fusion module lowers most metrics, implying each component contributes to the reported quality."],"supporting_citations":[{"why":"Supplies the denoising diffusion probabilistic model that trains the Stage I UNet.","marker":"[46]"},{"why":"Provides the UNet architecture used as the diffusion reconstruction network.","marker":"[53]"},{"why":"The prior diffusion-based fusion approach that avoids iterative sampling; DM-FNet extends it from infrared-visible to medical fusion.","marker":"[39]"},{"why":"Provides the AMFF attention module used for cross-modal feature fusion.","marker":"[54]"},{"why":"Supplies the maximum-local-standard-deviation SSIM loss used in the hybrid loss.","marker":"[55]"},{"why":"A two-stage autoencoder fusion baseline whose unified-fusion goal DM-FNet builds on.","marker":"[27]"},{"why":"A unified fusion baseline that still requires task-specific parameters, serving as the contrast case for DM-FNet's single parameter set.","marker":"[41]"},{"why":"A single-task multimodal transformer baseline for the MRI-PET and MRI-SPECT comparisons.","marker":"[1]"}],"fun_headline_variants":["Diffusion-trained encoder-decoder unifies MRI-CT, PET, SPECT fusion","One diffusion-trained network beats nine fusion methods on most metrics","DM-FNet: diffusion training powers a single fusion model across modalities","Diffusion process trains UNet that fuses medical images with top metrics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's central assumption is that feeding the fusion network source images corrupted with Gaussian noise at time steps 5, 10, and 20 produces features from the diffusion-trained UNet that are genuinely more informative than features from clean inputs or from a conventionally trained encoder.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion-trained encoder-decoder unifies MRI-CT, PET, SPECT fusion","One diffusion-trained network beats nine fusion methods on most metrics","DM-FNet: diffusion training powers a single fusion model across modalities","Diffusion process trains UNet that fuses medical images with top metrics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000768,"raw_usage":{"total_tokens":3467,"prompt_tokens":1072,"completion_tokens":2395,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":688,"completion_tokens_details":{"reasoning_tokens":2318}},"tokens_in":688,"tokens_out":2395,"duration_ms":19785,"temperature":1.0,"reasoning_tokens":2318,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:40:20.938658+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the Stage I encoder with the same UNet and a plain reconstruction loss (no diffusion schedule) while keeping Stage II identical, and compare on the three test sets; if the non-diffusion encoder matches or exceeds DM-FNet on the nine metrics, the diffusion process itself is not the load-bearing component.","supporting_citations":[{"cited_title":"Dea-net: Single image dehazing based on detail-enhanced convolution and content-guided attention.IEEE Transactions on Image Processing, 33:1002–1015, 2023","cited_arxiv_id":null,"evidence_quote":"Provides the AMFF attention module used for cross-modal feature fusion."},{"cited_title":"Fleet, and Mohammad Norouzi","cited_arxiv_id":null,"evidence_quote":"Provides the UNet architecture used as the diffusion reconstruction network."},{"cited_title":"Dif-fusion: Toward high color fidelity in infrared and visible image fusion with diffusion models.IEEE Transactions on Image Processing, 32:5705– 5720, 2023","cited_arxiv_id":null,"evidence_quote":"The prior diffusion-based fusion approach that avoids iterative sampling; DM-FNet extends it from infrared-visible to medical fusion."},{"cited_title":"M4fnet: Multimodal medical image fusion network via multi-receptive-field and multi-scale feature integration.Computers in biology and medicine, 159:106923, 2023","cited_arxiv_id":null,"evidence_quote":"Supplies the maximum-local-standard-deviation SSIM loss used in the hybrid loss."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A two-stage autoencoder fusion baseline whose unified-fusion goal DM-FNet builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A unified fusion baseline that still requires task-specific parameters, serving as the contrast case for DM-FNet's single parameter set."},{"cited_title":"Liu, and Yansong Duan","cited_arxiv_id":null,"evidence_quote":"A single-task multimodal transformer baseline for the MRI-PET and MRI-SPECT comparisons."}],"review_version":2}