{"id":"f2e26d5f-fc15-4475-a403-4774044b78b0","arxiv_id":"2504.16958","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ICONet, an iterative collaboration network that fuses reconstruction and super-resolution branches with Mamba-based spatial-channel feature learning, achieves the best PSNR and SSIM on the IXI, fastMRI, and Hecktor benchmarks among the compared methods.","lead":"A new deep learning network, ICONet, improves medical image super-resolution by alternating between an artifact-removal step and an upscaling step, guided by a Mamba-based feature extractor. The paper reports small but consistent PSNR and SSIM gains over existing methods on MRI and PET benchmarks, plus a small gain on a downstream tumor segmentation test.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Empirical support for 'outperforms most SOTA' rests on a small, partly dated baseline set with sub-0.15 dB margins on fastMRI and no error bars; this should be tested against current Mamba/Transformer SR methods under matched training.","rationale":"The reader's weakest_assumption was that the synthetic degradation model may not transfer to real acquisitions. That is a legitimate external-validity concern, but it would only matter after the benchmark superiority is established. My stress-test lands one step earlier: the benchmark comparison itself is not yet strong enough to support 'outperforms most state-of-the-art approaches.' The main tables omit recent strong baselines, and the fastMRI margins are tiny. Because the paper provides no code, no seed information, and no error bars, I cannot tell whether the reported gains are signal or noise. This is not a claim of misconduct; it is a request for the minimal evidence needed to make the headline claim falsifiable. The ablation tables (IV, V, VI, VIII) are internally consistent and support the iterative-prior design as a useful contribution; the issue is specifically the external comparison supporting the headline superiority. My recommended verdict remains CONDITIONAL: the paper is promising but should not be accepted as demonstrating SOTA superiority until a matched comparison against current Mamba/Transformer SR methods with uncertainty estimates is provided. Hence verdict_should_be is UNCHANGED relative to the reader's conditional decision, with the condition made more specific.","tokens_in":16793,"tokens_out":8162,"duration_ms":81174,"concrete_test":"Obtain or re-implement SwinIR [51] and Deform-Mamba [41] (plus DiVANet/T2Net for control), train them on the exact same IXI/fastMRI splits with the same Ref. [50] degradation, optimizer, and training schedule used for ICONet, and run 5 seeds per method. Report mean±std PSNR/SSIM for 2x and 4x plus paired bootstrap 95% CIs on the ICONet-minus-baseline differences. If any fastMRI CI includes 0 or a Mamba/Transformer baseline equals/beats ICONet, the central claim in the abstract needs to be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is quantitative: Tables I-III show ICONet with the best PSNR/SSIM on IXI, fastMRI, and Hecktor. The load-bearing condition is that those comparisons are against an adequate set of current state-of-the-art methods under matched, repeatable conditions. That condition is not met. Section IV-C compares only SRCNN, VDSR, FMISR, T2Net, and DiVANet; SwinIR appears only in the Table IX module-complexity ablation, and no recent Mamba-based SR model (e.g., the authors' own Deform-Mamba, Ref. [41]) is in the main tables. On fastMRI, the reported advantage is 0.12 dB (2x) and 0.13 dB (4x) over T2Net, and 0.14/0.07 dB over DiVANet; no standard deviations, seed counts, or paired tests are reported, so these margins are within the range of normal run-to-run variation. The claim therefore rests on point estimates from a narrow baseline set; a single seed or a missing strong baseline could change the ranking. The design idea is not contradicted, but the headline superiority claim is under-supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ICONet, an iterative collaboration network for medical image super-resolution that combines a reconstruction branch, a super-resolution branch, and an SR-Rec fusion module. The reconstruction branch generates an artifact-free image that guides the super-resolution branch, and the two branches interact iteratively across multiple upsampling stages. The feature extraction backbone is a residual spatial-channel feature learning (RSCFL) module built on vision Mamba blocks and a channel ranking block. Experiments are reported on IXI (brain MRI), fastMRI (knee MRI), and Hecktor (PET) datasets under 2x and 4x upsampling, plus a downstream brain tumor segmentation task. The paper reports the best PSNR/SSIM among the compared methods and provides ablations showing contributions from the reconstruction branch, fusion module, loss weight, number of blocks, and multi-stage supervision.","tokens_in":17087,"tokens_out":6009,"duration_ms":56176,"significance":"If the reported results hold, ICONet would be a meaningful contribution to medical image super-resolution by demonstrating that iterative multi-task collaboration with a Mamba-based backbone can improve both performance and efficiency relative to Transformer-based alternatives. The design idea of progressively propagating reconstruction priors into super-resolution is clearly articulated and the ablation study gives initial evidence that each component contributes. The inclusion of a downstream segmentation evaluation is a positive step toward clinical relevance. However, the strength of the headline claim is currently limited by the narrow baseline set, the absence of statistical rigor, and the reliance on a single synthetic degradation model, which together prevent the paper from fully supporting its claim of outperforming state-of-the-art approaches.","major_comments":[{"comment":"The state-of-the-art comparison includes only SRCNN, VDSR, FMISR, T2Net, and DiVANet. Recent medical SR methods (e.g., TransMRSR [13], MFTN [33], Deform-Mamba [41], and the authors' own self-prior guided Mamba network [42]) are omitted from the quantitative tables, even though they are cited in the related work. Furthermore, no error bars, seeds, or significance tests are reported. On fastMRI, the margins over T2Net/DiVANet are 0.12-0.14 dB at 2x and 0.07-0.14 dB at 4x, which are within the range of typical run-to-run variation. The claim that ICONet outperforms most state-of-the-art approaches is therefore not established. Please add the missing baselines, at least three seeds with mean and standard deviation, and appropriate statistical tests.","section":"IV-C, Tables I-III"},{"comment":"Training details are insufficient for reproducibility. The authors report the optimizer, learning rate, feature dimension, and loss weights, but omit the number of training epochs, batch size, patch size, data augmentation, learning rate schedule, and the exact procedure for generating low-resolution images with the frequency-domain degradation model of Ref. [50]. Without these details (and ideally code release), the experimental numbers cannot be independently reproduced. Please provide a complete training configuration and make the code publicly available.","section":"IV-B"},{"comment":"All experiments use a single synthetic degradation model from Ref. [50]. The statement that this model 'helps generate LR images that more closely resemble the characteristics of real-world scenarios' is not supported by any experiment. Because the method is intended for clinical imaging, the authors should evaluate at least one additional degradation model (e.g., k-space undersampling with noise for MRI) or a real paired LR/HR dataset to demonstrate that the reported gains are not an artifact of the specific synthetic degradation.","section":"IV-B and evaluation setup"},{"comment":"The comparison with SwinIR appears to use a non-standard configuration: 1.260M parameters and 20.66 GFLOPs, whereas standard SwinIR typically has about 11.5M parameters. This may put SwinIR at a disadvantage. The authors should specify the exact SwinIR configuration, use a recognized standard baseline, and also compare with other Transformer-based SR models (e.g., RCAN or SwinIR-light) to substantiate the efficiency claim.","section":"IV-D6, Table IX"},{"comment":"The downstream segmentation experiment reports a single Dice score for each method without variance or significance testing. The setup also includes high-resolution T1 images as auxiliary inputs, but the segmentation training protocol is not described in detail. This makes it difficult to attribute the improvement to the SR method rather than to random variation or the additional T1 input. Please report multiple runs, standard deviations, and a statistical test, and clarify the segmentation training procedure.","section":"IV-E, Table X"}],"minor_comments":[{"comment":"There is a placeholder citation in the text 'traditional super-resolution approaches [ ?], [12]'; the question mark should be replaced with the appropriate reference.","section":"Introduction, Section I"},{"comment":"The degradation model from Ref. [50] is referenced but not described; please summarize the model and how it is applied to each dataset, especially for PET images where the acquisition physics differs from MRI.","section":"Section IV-B"},{"comment":"The notation for the semantic attention block is dense; in particular, the symmetric computation for F_Rec_SAB is mentioned in the text but not written out. For clarity, define both directions explicitly or state that they follow the same form.","section":"Section III-C, Eq. (4)-(7)"},{"comment":"The phrase 'black box-like sharable features' is vague; please provide a more precise explanation of why parallel multi-task sharing underperforms the serial framework.","section":"Section IV-D2"},{"comment":"The abstract and Section I say ICONet 'outperforms most state-of-the-art approaches,' while the conclusion says it 'greatly outperforms state-of-the-art medical image super-resolution approaches.' This inconsistency should be resolved after the comparison set is expanded.","section":"Section IV-C and Conclusion"},{"comment":"The reconstruction branch is trained on the same ground-truth images as the super-resolution branch, so the resulting 'prior' is a learned self-prior rather than an external or physics-based constraint. The paper should explicitly acknowledge this and discuss its implications for the method's generalizability.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is from a group that has published closely related Mamba-based MRI super-resolution work (Refs. [41], [42]). It is surprising that these methods, especially Deform-Mamba, are cited in the related work but not included in the quantitative comparisons; this omission weakens the novelty and superiority claims. The manuscript also contains a placeholder citation in the Introduction, which suggests the paper may have been submitted before thorough proofreading. Given that the central design idea is coherent and the ablations support internal consistency, major revision is appropriate if the comparison and reproducibility issues are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I've read the ICONet paper. Short version: the architecture is a genuine contribution, but the headline superiority claim is over-extended by a thin baseline set and point estimates. I'd send it to review, with the expectation that the authors get pushed on both.\n\nWhat is new: ICONet's iterative collaboration between reconstruction and SR branches, with SR output fed back into reconstruction across stages, is a specific arrangement not in the prior work I know. The RSCFL module (vision Mamba + channel ranking) is a reasonable combination, and the complexity ablation against SwinIR makes a credible efficiency case. The experiments span three datasets and include a downstream BraTS segmentation check, which is a nice practical touch. Ablations are standard and show each component matters.\n\nThe soft spots are real but not fatal. The baseline set is short and dated: SRCNN, VDSR, FMISR, T2Net, DiVANet. The authors cite Deform-Mamba and other Mamba-based SR work but don't compare against any of it. That would be more acceptable if margins were large; they aren't. On fastMRI 4x, ICONet beats DiVANet by 0.07 dB and T2Net by 0.13 dB, with no error bars, seed counts, or significance tests. Those gaps are within run-to-run variance. On IXI the gains are larger (0.6-0.8 dB) and more convincing. The degradation model is synthetic; the authors only say it 'helps generate LR images that more closely resemble' real scenarios, which isn't verifiable without code. Code would fix a lot.\n\nOn the circularity point: the reconstruction prior is internally generated, but that's a self-prior, not an external constraint. That's legitimate; the paper just shouldn't imply it's an external prior. The concluding limitation about not handling frequency data is honest, though minor.\n\nWho this is for: anyone working on Mamba-based medical SR or joint reconstruction/SR. It's not a paradigm shift, but the iterative collaboration framing is useful and the downstream segmentation result shows the improvements propagate. I'd bring it to a reading group as a representative multi-task SR paper with a state-space backbone. It deserves a serious referee: the idea is specific, experiments are internally consistent, and despite my skepticism about the magnitude claim, the work isn't sloppy. My recommendation: send to peer review. The review should require a wider set of current baselines (at least Deform-Mamba and a recent Transformer like MFTN) and variance estimates, or a clear declaration that numbers come from a single run. Then the central claim becomes testable.","headline":"A real architecture in need of stronger baselines and error bars before the superiority claim holds up.","tokens_in":17586,"tokens_out":3554,"would_cite":true,"duration_ms":34050,"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":"The paper claims that medical image super-resolution improves when an artifact-cleaning reconstruction branch and an upscaling branch exchange information iteratively, with every intermediate stage supervised.","keywords":["medical image super-resolution","multi-task learning","reconstruction prior","iterative collaboration network","vision Mamba","channel attention","MRI","PET"],"falsifier":"Construct low-resolution medical images through a different realistic degradation (for example, Gaussian blur plus noise, or undersampled k-space reconstruction), and retrain the same network; if its PSNR/SSIM advantage over T2Net disappears or reverses, the iterative reconstruction-prior claim is specific to the synthetic degradation rather than general.","tokens_in":16651,"feed_emoji":"🧠","tokens_out":6814,"duration_ms":59493,"temperature":0.7,"pith_summary":"The paper sets out to show that medical image super-resolution improves when the network is organized as an iterative collaboration between two tasks: a reconstruction branch first produces an artifact-free version of the low-resolution input, and a super-resolution branch uses that cleaned image as a prior while upscaling. The two branches exchange features through a fusion module, and the intermediate super-resolved image is fed back into the reconstruction branch for the next stage, with every stage supervised against its own ground truth. The authors report that this loop outperforms single-task super-resolution networks and a representative parallel multi-task network on brain MRI, knee MRI, and PET data at both 2x and 4x magnification, and that the resulting images support better downstream brain-tumor segmentation. The significance, if the claim holds, is that reconstruction can serve as an internal, reference-free prior that progressively guides super-resolution, rather than requiring external priors.","feed_headline":"Reconstruction loop lifts medical image super-resolution","feed_subtitle":"Iterative two-branch design tops PSNR/SSIM tables on MRI and PET at 2x and 4x upscaling.","key_machinery":"The load-bearing mechanism is the iterative collaboration loop itself: reconstruction output feeds super-resolution, and super-resolved output feeds the next reconstruction stage, so the two tasks refine each other. Inside each branch, the residual spatial-channel feature learning (RSCFL) module combines a vision Mamba block, which scans image patches in four directions with a linear-complexity selective state-space model to capture long-range spatial dependencies, with a channel ranking block, which reweights feature channels to suppress background noise. The SR-Rec fusion module computes semantic attention scores between SR and reconstruction features and adaptively combines them, so reconstruction information is transferred where it is useful. Multi-stage L1 losses supervise every intermediate output, decomposing a 4x upsampling into two supervised 2x steps.","core_discovery":"In the paper's own terms, the discovery is that reconstruction prior, iteratively fused into the super-resolution branch, progressively improves image quality more than either single-task super-resolution or parallel multi-task learning. The reconstruction branch generates an artifact-free image Rec1 from LR; the SR branch upsamples it to SR2x; the SR-Rec fusion module transfers reconstruction features into the SR branch via semantic attention; then SR2x is re-fed into the reconstruction branch to produce Rec2, which guides the final SR4x. Multi-stage L1 supervision on each reconstruction and SR output stabilizes this loop. The reported quantitative tables show the best PSNR/SSIM among the compared methods on IXI and fastMRI at 2x and 4x, and on Hecktor against T2Net, and the ablation study attributes the gains to the reconstruction branch, the fusion module, and multi-stage supervision.","pith_inferences":["A natural extension, not tested in the paper, is whether the loop still wins when low-resolution images come from real accelerated acquisitions or other degradation kernels; all benchmark gains are measured on images synthesized with one frequency-domain degradation model.","The same iterative clean-then-upscale loop could transfer to other restoration pairs, such as denoising plus super-resolution or inpainting plus super-resolution, where one branch removes a corruption and the other enlarges.","The downstream segmentation experiment suggests that task-based evaluation could be more clinically meaningful than PSNR/SSIM alone, though the reported Dice gap is small and measured in one segmentation setting."],"forward_implications":["If the central claim holds, serial coupling of reconstruction and super-resolution should be preferred over parallel shared-feature coupling; the paper's ablation reports 33.76 vs 33.41 PSNR at 2x on IXI.","Linear-complexity state-space feature extraction can replace Transformers for medical super-resolution at lower parameter and FLOP counts; the paper reports 0.996M params and 9.61G FLOPs for vision Mamba vs 1.260M and 20.66G for SwinIR at similar or better PSNR.","Decomposing high upsampling into supervised 2x stages reduces error accumulation; multi-stage supervision improves 4x PSNR from 31.02 to 31.24 dB over single-stage supervision in the ablation.","Super-resolved images produced under reconstruction guidance improve downstream tumor segmentation; the paper reports Dice 0.5265 vs 0.5019 with T2Net-generated images."],"supporting_citations":[{"why":"Supplies the prior parallel multi-task reconstruction-plus-SR architecture that ICONet is compared against and extends.","marker":"[8]"},{"why":"Supplies the frequency-domain degradation model used to generate the low-resolution training and test images.","marker":"[50]"},{"why":"Baseline single-task CNN super-resolution method in the quantitative comparisons.","marker":"[4]"},{"why":"Baseline very deep CNN super-resolution method in the quantitative comparisons.","marker":"[5]"},{"why":"Baseline fast medical image super-resolution method in the quantitative comparisons.","marker":"[31]"},{"why":"Baseline attention-based super-resolution method in the quantitative comparisons.","marker":"[49]"},{"why":"Supplies the selective state-space model that underlies the vision Mamba block in RSCFL.","marker":"[18]"},{"why":"Supplies the channel-attention mechanism that the channel ranking block is inspired by.","marker":"[43]"},{"why":"SwinIR serves as the Transformer-based super-resolution baseline for the efficiency and performance comparison in Table IX.","marker":"[51]"},{"why":"Attention U-Net is the backbone used in the downstream brain-tumor segmentation evaluation.","marker":"[53]"}],"fun_headline_variants":["Iterative reconstruction prior boosts medical image SR","Collaborative loop sharpens medical images","Reconstruction-guided iterations refine MRI and PET","Two-branch iteration bests single-task upscaling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains depend on the low-resolution images being produced by a specific synthetic frequency-domain degradation model; if that model does not resemble real clinical acquisition, the improvements may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Iterative reconstruction prior boosts medical image SR","Collaborative loop sharpens medical images","Reconstruction-guided iterations refine MRI and PET","Two-branch iteration bests single-task upscaling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000627,"raw_usage":{"total_tokens":2923,"prompt_tokens":993,"completion_tokens":1930,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":609,"completion_tokens_details":{"reasoning_tokens":1873}},"tokens_in":609,"tokens_out":1930,"duration_ms":14398,"temperature":1.0,"reasoning_tokens":1873,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:04:15.458424+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct low-resolution medical images through a different realistic degradation (for example, Gaussian blur plus noise, or undersampled k-space reconstruction), and retrain the same network; if its PSNR/SSIM advantage over T2Net disappears or reverses, the iterative reconstruction-prior claim is specific to the synthetic degradation rather than general.","supporting_citations":[{"cited_title":"Task transformer network for joint MRI reconstruction and super-resolution,","cited_arxiv_id":null,"evidence_quote":"Supplies the prior parallel multi-task reconstruction-plus-SR architecture that ICONet is compared against and extends."},{"cited_title":"Multi-contrast super-resolution MRI through a progressive network,","cited_arxiv_id":null,"evidence_quote":"Supplies the frequency-domain degradation model used to generate the low-resolution training and test images."},{"cited_title":"Image super-resolution using deep convolutional networks,","cited_arxiv_id":null,"evidence_quote":"Baseline single-task CNN super-resolution method in the quantitative comparisons."},{"cited_title":"Accurate image super-resolution using very deep convolutional networks,","cited_arxiv_id":null,"evidence_quote":"Baseline very deep CNN super-resolution method in the quantitative comparisons."},{"cited_title":"A fast medical image super resolution method based on deep learning network,","cited_arxiv_id":null,"evidence_quote":"Baseline fast medical image super-resolution method in the quantitative comparisons."},{"cited_title":"Single image super-resolution based on directional vari- ance attention network,","cited_arxiv_id":null,"evidence_quote":"Baseline attention-based super-resolution method in the quantitative comparisons."},{"cited_title":"Squeeze-and-excitation networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the channel-attention mechanism that the channel ranking block is inspired by."},{"cited_title":"Swinir: Image restoration using swin transformer,","cited_arxiv_id":null,"evidence_quote":"SwinIR serves as the Transformer-based super-resolution baseline for the efficiency and performance comparison in Table IX."}],"review_version":1}