{"id":"cfb400c5-e362-4299-9075-1b5eb4a56884","arxiv_id":"2505.14715","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":1.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A broad literature review of multi-modal medical image fusion techniques, applications, and challenges, with no new experimental results and several technical errors.","lead":"This paper surveys the field of multi-modal medical image fusion, which combines CT, MRI, PET, SPECT, and ultrasound images to improve diagnosis. It organizes methods, algorithms, clinical applications, and challenges into a single reference for clinicians and researchers entering the area.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract's claim that MMIF 'significantly advance[s] diagnostic accuracy' rests on image-quality metrics that the review itself concedes are not yet validated against clinical outcomes; Section 9.5 undercuts the positive synthesis.","rationale":"The reader's weakest assumption—that quantitative fusion metrics are treated as proxies for clinical benefit—is exactly the gap we identify. The review itself admits in Section 9.5 that rigorous multi-center validation is lacking and that fused-image errors can cause misdiagnosis, which contradicts the categorical claim in the Abstract that MMIF 'significantly advance[s] diagnostic accuracy.' We do not see a need to change the verdict: the paper can still serve as a broad survey if the authors disclose selection methodology and temper the clinical claims, which is what CONDITIONAL already asks for. We did not find grounds for REJECT because the central claim is plausible and the review contains a useful compendium of techniques and applications; the issue is one of overstatement and unverified proxy metrics, not fraud. We also note the metric-definition errors as corroborating evidence that the paper's quantitative basis has not been rigorously checked, but the core problem is the missing link between image-quality metrics and clinical outcomes.","tokens_in":43323,"tokens_out":4358,"duration_ms":46186,"concrete_test":"Perform a systematic audit of every study cited in Table 7 and in Sections 6–8: classify each by whether it reports (a) only fused-image quality metrics (SSIM, PSNR, entropy, etc.) or (b) a clinical diagnostic endpoint (reader study, sensitivity/specificity/AUC, or patient outcome). If fewer than 20% of the surveyed studies include any clinical endpoint, then the Abstract's claim of 'significantly advance diagnostic accuracy' is an extrapolation, and the verdict should require the authors to temper that claim. As a secondary check, re-derive Eq. (8) against the standard SSIM definition (Wang et al., 2004) and verify the two Sobel kernels in Eq. (3); if either is wrong, the review's metric-based rankings need re-examination.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, stated in the Abstract and Conclusion, is that MMIF 'significantly advance[s] diagnostic accuracy, lesion detection, and segmentation' and is poised to transform clinical care. The evidence marshaled for this claim consists almost entirely of studies reporting fused-image quality metrics such as SSIM, PSNR, entropy, average gradient, and spatial frequency (Section 7, Tables 6–7). Nowhere does the review establish that improvements in these metrics translate into improved diagnostic accuracy, lesion detection, or patient outcomes. The paper's own Section 9.5 concedes the point: 'Advanced image fusion techniques require rigorous validation through costly and time-consuming multi-center trials,' errors in fused images 'can carry serious consequences, including misdiagnosis,' and such validation 'is still lacking.' This creates a direct internal tension: the Abstract asserts a significant clinical benefit, but the evidence section admits the field has not yet met the burden of proof. The load-bearing assumption is that quantitative fusion metrics are reliable proxies for clinical diagnostic benefit. If that assumption fails, the review's positive synthesis overstates MMIF's clinical readiness. Two supporting technical errors reinforce this concern: Eq. (8) defines SSIM asymmetrically and not bounded by 1 (standard SSIM is symmetric and ≤1), and the two Sobel kernels in Eq. (3) are identical, making the edge-intensity measure degenerate. These errors suggest that the quantitative comparisons the review relies on have not been carefully checked, further weakening confidence in metric-based conclusions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a narrative survey of multi-modal medical image fusion (MMIF), covering the major imaging modalities (X-ray, CT, MRI, ultrasound, fMRI, PET, SPECT), the standard taxonomy of fusion levels (pixel-, feature-, hybrid, and decision-level), a catalogue of fusion algorithms (morphological, HVS-based, neural-network, sub-band, and fuzzy-logic methods), evaluation metrics, clinical applications, challenges, and future directions. The paper claims that MMIF 'significantly advance[s] diagnostic accuracy, lesion detection, and segmentation' and positions the field as poised for routine clinical adoption. The review is organized as a broad reference rather than a focused technical contribution, with Tables 6 and 7 providing a large compilation of recent studies, fusion techniques, modalities, and reported contributions.","tokens_in":43573,"tokens_out":3916,"duration_ms":40755,"significance":"If the survey's technical content were accurate and its clinical claims appropriately calibrated, this would be a useful reference for researchers and clinicians entering the MMIF area. The paper's strengths are its breadth of coverage, its structured presentation of fusion levels and algorithm families, and its explicit enumeration of challenges such as data privacy, standardization, and the need for clinical validation. The authors also deserve credit for acknowledging, in Section 9.5, that rigorous multi-center clinical validation is still lacking. However, the review's positive synthesis is undermined by (i) two concrete technical errors in the definitions of standard metrics that are used as evidence throughout the paper, and (ii) a mismatch between the abstract's strong clinical-readiness claims and the paper's own admission that fusion metrics have not been shown to translate into improved patient outcomes. These issues are fixable, but they currently limit the manuscript's reliability as a reference.","major_comments":[{"comment":"The SSIM formula as written is inverted relative to the standard definition. The conventional expression is SSIM(x,y) = [(2μxμy + C1)(2σxy + C2)] / [(μx² + μy² + C1)(σx² + σy² + C2)], giving values bounded by 1, but Eq. (8) has the numerator and denominator swapped. As written, the stated property that 'SSIM values close to 1 indicate high structural similarity' is not guaranteed, and the metric can exceed 1. Since SSIM is a central evaluation metric referenced throughout the survey, this error needs correction.","section":"Section 7.8, Eq. (8)"},{"comment":"The two Sobel kernels displayed for hx and hy are identical, and neither matches the standard Sobel operator: the x-kernel has the wrong sign in the first and third rows, and the y-kernel is missing. The edge intensity measure EI is therefore degenerate, as it cannot distinguish horizontal and vertical gradient information. This undermines the credibility of edge-intensity as an evaluation metric in the survey's technical presentation.","section":"Section 7.3, Eq. (3)"},{"comment":"The Abstract and Conclusion state that MMIF 'significantly advance[s] diagnostic accuracy, lesion detection, and segmentation' and that the field is close to 'routine clinical use,' but Section 9.5 concedes that 'rigorous multi-center clinical validation is still lacking' and that errors in fused images 'can carry serious consequences, including misdiagnosis.' The evidence cited in Tables 6–7 consists mainly of image-quality metrics (SSIM, PSNR, entropy, average gradient, spatial frequency) with no established connection to patient-level diagnostic outcomes. The manuscript should either temper its clinical-readiness claims to match the stated evidence or provide a reasoned justification for why these quality metrics are valid proxies for clinical benefit.","section":"Abstract and Conclusion vs. Section 9.5"}],"minor_comments":[{"comment":"The subsection titled 'Decision-learning-based fusion' actually describes dictionary-learning-based fusion (e.g., Gaussian filter with integrated dictionary learning, image patch sampling, sparse representation), which is a feature-level technique, not a decision-level one. This placement is inconsistent with the taxonomy in Section 4 and with Table 4, and it creates confusion about the paper's classification scheme.","section":"Section 4.4.1"},{"comment":"The heading 'Algorithms using Human Value System operations' should be 'Human Visual System operations' to match the text's own abbreviation HVS.","section":"Section 5.2 heading"},{"comment":"Several typos should be corrected: 'prons' should be 'pros' in Tables 1 and 2, and 'indivisdual sternghts, limistations' should be 'individual strengths, limitations' in Table 4.","section":"Tables 1, 2, and 4"},{"comment":"There are repeated grammatical and spelling issues throughout the manuscript (e.g., 'In contrary to CT' in Section 2.1.3, 'where RF row frequency and CF (column frequency) are computed' in Section 7.7). A careful proofreading pass is needed.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper cites the authors' own prior works in several places (e.g., refs 13, 16, 19–22, 24), but the survey's conclusions do not depend on those works; the self-citation rate is within a normal range for an active research group. The manuscript would benefit from a dedicated technical proofread by the authors before resubmission, with particular attention to Eqs. (3) and (8) and to aligning the abstract's claims with the limitations stated in Section 9.5."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi colleague,\n\nQuick take on arXiv:2505.14715, the MMIF review by Zubair et al. It's a review, so don't expect new results. What it does well is organize the field: the taxonomy of pixel-, feature-, hybrid-, and decision-level fusion, and the big tables mapping techniques to modalities and clinical applications (Tables 6 and 7) are a genuinely useful entry point for someone new to MMIF. The narrative is broadly consistent with the literature I know, and the challenge sections (9 and 10) hit the right topics: computational cost, data privacy, heterogeneity, interpretability, and the gap between algorithmic development and clinical adoption.\n\nThe soft spots are real but fixable. First, the equations. The SSIM formula (Eq. 8) is wrong—it inverts numerator and denominator and isn't bounded by 1. The two Sobel kernels in Eq. 3 are identical, which makes the edge-intensity computation degenerate. Those aren't just typos; they're in the section that's supposed to define the metrics the review leans on. Second, there's no stated methodology for literature selection. It calls itself 'comprehensive' but doesn't say how references were chosen, so its authority as a systematic map is limited. Third, and most important, the clinical claim outruns the evidence. The abstract says MMIF 'significantly advances diagnostic accuracy, lesion detection, and segmentation,' but the evidence cited is mostly image-quality metrics—SSIM, PSNR, entropy. The paper's own Section 9.5 admits that rigorous multi-center clinical validation is lacking and that errors in fused images can cause misdiagnosis. That's an internal tension that needs resolving: either the abstract gets tempered or the evidence section gets strengthened.\n\nSelf-citations are present but not load-bearing, so I wouldn't hold that against it.\n\nMy recommendation: if this lands on an editor's desk, it merits peer review rather than a desk reject. A competent referee would ask for major revision: correct the equations, add a search strategy, and bring the abstract and conclusion in line with the actually-available evidence. I'd use the paper as a starting point for a reading group discussion on how image-fusion evaluation metrics relate (or don't) to clinical outcomes—that's a valuable conversation. I wouldn't cite it in my own work because I'd rather cite the primary sources.\n\nCheers.","headline":"Useful organizational survey of MMIF, but with wrong equations and clinical claims that outrun the evidence.","tokens_in":44107,"tokens_out":3059,"would_cite":false,"duration_ms":32405,"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":"This review claims that fusing structural and functional medical scans into one image improves diagnostic accuracy, and maps the techniques, applications, and obstacles behind that claim.","keywords":["multi-modal medical image fusion","medical imaging modalities","image fusion algorithms","deep learning fusion","transformer-based fusion","computer-aided diagnosis","clinical decision support","diagnostic accuracy"],"falsifier":"A prospective, multi-center reader study in which blinded clinicians diagnose the same cases from fused images and from the best single modality would settle the central claim: if fused reads do not improve diagnostic accuracy, lesion detection, or inter-reader agreement over the single best modality, the review's synthesis overstates MMIF's clinical readiness. A cheaper partial test is correlational: compute whether the SSIM, PSNR, and entropy gains reported across the surveyed papers predict expert-rated diagnostic utility on the same image pairs.","tokens_in":43118,"feed_emoji":"🩻","tokens_out":7435,"duration_ms":71637,"temperature":0.7,"pith_summary":"This paper is a systematic survey arguing that multi-modal medical image fusion (MMIF), which combines structural scans like CT and MRI with functional scans like PET and SPECT into a single image, improves diagnostic accuracy, lesion detection, and treatment planning. Its contribution is reference value: a structured map of the fusion levels, the algorithm families used at each level, the modality combinations applied to specific organs, and the clinical tasks each combination serves. The authors take the field's progress as real, from hand-crafted pixel-level rules to deep learning and transformer architectures, but they stress that routine clinical adoption still awaits explainable AI, privacy-preserving training, real-time systems, and standardization.","feed_headline":"Merging CT, MRI, and PET scans into a single image aids diagnosis","feed_subtitle":"Survey maps the algorithms and clinical applications behind image fusion, and flags the missing clinical proof.","key_machinery":"The load-bearing structure is the review's taxonomy: a three-level hierarchy of fusion abstraction, pixel-level, feature-level, and decision-level, with hybrid pixel-feature methods in between, cross-referenced against six algorithm families (morphological operations, human-visual-system operators, neural networks, sub-band decomposition, fuzzy logic, and hybrid methods) and against tables mapping modality combinations to organs. These tables carry the argument: they are what allow the review to claim that specific techniques, such as NSST with pulse-coupled neural networks for MRI-PET or guided filtering for CT-MRI, fit specific clinical tasks. The accompanying evaluation metrics, SSIM, PSNR, entropy, spatial frequency, and the Xydeas-Petrovic edge-preservation index, supply the quantitative language in which the surveyed papers report success.","core_discovery":"The review's central claim is that no single imaging modality captures the full complexity of disease, and that fusing complementary modalities yields a representation genuinely more informative than any input alone: CT's anatomy plus PET's metabolism, or MRI's soft-tissue contrast plus SPECT's perfusion data. On that basis it asserts that MMIF significantly advances diagnostic accuracy, lesion detection, and segmentation, and it treats PET/CT in oncology as the mature proof of concept. The review's own reading of the literature is that the field has moved through three generations, from hand-crafted pixel- and transform-domain rules, to intelligent optimization and hybrid methods, to deep learning, GAN, and transformer models that learn fusion end to end, with each generation shifting the trade-off between robustness, computational cost, and interpretability. It also concludes that clinical use is still gated by missing standardized datasets, computational demands, privacy regulation, and the absence of rigorous multi-center validation.","pith_inferences":["A reader could take the abstract's confident framing as a research agenda rather than an established clinical verdict: the review's own Section 9.5 concedes that rigorous clinical validation is lacking, so the strongest defensible claim is that MMIF is promising and well mapped, not yet proven at the bedside.","The proxy assumption that SSIM, PSNR, and entropy gains equal diagnostic benefit is directly testable: one could correlate the published fusion-metric improvements with expert radiologist ratings or with measured changes in diagnostic accuracy on the same image pairs.","The taxonomy suggests a prediction the authors do not draw: decision-level fusion, currently the least-used level, should gain ground as imaging is fused with genomic, proteomic, and clinical data, since decision-level methods are the natural fit for heterogeneous non-image inputs.","An editorial test of the survey's organizing claim would be a reader study comparing fused images against the best single modality for the same cases, because the review assembles little direct evidence that fused images change clinician decisions."],"forward_implications":["If the review is right, hybrid PET/CT-style fusion will extend to more organ systems and modality pairs, and PET/MRI will keep expanding in neurology and oncology as hardware matures.","Deep learning and transformer-based fusion, which learn cross-modal mappings without hand-crafted rules, will keep displacing classical pixel-level and transform-domain methods as the default research direction.","Clinical adoption will hinge on the bottlenecks the review names: explainable models clinicians can trust, real-time point-of-care systems, federated learning for privacy, and regulatory standardization.","Fusion quality will have to be judged by task-level outcomes such as segmentation accuracy, diagnostic confidence, and decision impact rather than by pixel-level metrics alone, a shift the review explicitly endorses.","The review's negative finding, that standardized multi-modal datasets and multi-center validation are still scarce, implies that reported performance gains in the literature have not yet been converted into generalizable clinical tools."],"supporting_citations":[{"why":"Provides the survey's definitional basis for pixel-level fusion and its state-of-the-art taxonomy.","marker":"[114]"},{"why":"Earlier review of multimodality medical image fusion that grounds the paper's motivation and scope.","marker":"[3]"},{"why":"Historical account of PET/CT, the canonical fused modality pair the review treats as proof of concept.","marker":"[111]"},{"why":"Clinical application of PET/MRI in oncology, load-bearing for the review's claims of oncological and neurological benefit.","marker":"[100]"},{"why":"Reference textbook for PET physics and functional imaging principles underpinning the modality descriptions.","marker":"[71]"},{"why":"Representative recent fusion method using intuitionistic fuzzy sets, illustrating the algorithm families and metrics the review catalogues.","marker":"[96]"},{"why":"Transformer-based fusion model (MDC-RHT) cited as evidence of the deep-learning trend.","marker":"[93]"},{"why":"Federated-learning fusion framework that anchors the review's privacy-preserving future direction.","marker":"[154]"},{"why":"Multimodal deep learning combined with clinical and radiological data, supporting the claim that non-image data can be integrated into fusion pipelines.","marker":"[102]"}],"fun_headline_variants":["Fusing medical images: a review of methods and missing proof","Multi-modal image fusion review: techniques, challenges, clinical gaps","Combining CT, MRI, PET scans: survey of fusion methods and limits","Image fusion aids diagnosis, but clinical proof lags, review finds","How merging scans sharpens diagnosis: a comprehensive review"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The review's upbeat conclusion rests on treating the quantitative fusion metrics reported across surveyed studies (SSIM, PSNR, entropy) as proxies for real clinical benefit; the paper itself concedes in Section 9.5 that multi-center clinical validation is still absent and that errors in fused images can cause misdiagnosis.","fun_headline_variants_meta":{"raw":{"variants":["Fusing medical images: a review of methods and missing proof","Multi-modal image fusion review: techniques, challenges, clinical gaps","Combining CT, MRI, PET scans: survey of fusion methods and limits","Image fusion aids diagnosis, but clinical proof lags, review finds","How merging scans sharpens diagnosis: a comprehensive review"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000585,"raw_usage":{"total_tokens":2775,"prompt_tokens":994,"completion_tokens":1781,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":1693}},"tokens_in":610,"tokens_out":1781,"duration_ms":12659,"temperature":1.0,"reasoning_tokens":1693,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:30:56.663829+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A prospective, multi-center reader study in which blinded clinicians diagnose the same cases from fused images and from the best single modality would settle the central claim: if fused reads do not improve diagnostic accuracy, lesion detection, or inter-reader agreement over the single best modality, the review's synthesis overstates MMIF's clinical readiness. A cheaper partial test is correlational: compute whether the SSIM, PSNR, and entropy gains reported across the surveyed papers predict expert-rated diagnostic utility on the same image pairs.","supporting_citations":[],"review_version":1}