{"id":"29228ff6-e28b-4601-8f39-c938ae0717ea","arxiv_id":"2412.10452","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A dual-decoder CycleGAN with Cryosection segmentation fusion colorizes MRI scans while preserving structure, outperforming five baselines on the Visible Korean Human dataset.","lead":"This paper presents a deep learning method that colorizes MRI scans using colors and textures from Cryosection anatomy images while keeping the MRI structures intact. It combines segmentation information with a cycle-consistent GAN and reports better color and structure scores than five existing methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pixel-level FSIM/STSIM comparisons use misaligned Cryosection ground truth, contradicting the paper's own statement that deformations persist; this undermines the quantitative superiority claim.","rationale":"After reading the paper in good faith, I find the architecture and ablation study internally coherent: the dual-decoder design, compression-activation mechanism, and segmentation loss are plausible, and the ablation results degrade as expected when components are removed. The central claim, however, hinges on the quantitative comparison in Table I. The reader flagged train/test leakage from random slice splitting; that is a valid concern, but it affects all compared methods equally and would mainly inflate absolute numbers rather than necessarily change relative ranking. A more fundamental issue is the evaluation protocol: computing pixel-level FSIM/STSIM between the colorized MRI and the Cryosection reference assumes spatial correspondence that the paper itself says does not exist. This is not a disagreement with consensus; it is an internal inconsistency between the stated deformation problem and the chosen metrics. If the registration is imperfect, the reported FSIM/STSIM are contaminated by misalignment, and the claimed margin over baselines may be an artifact of how each method handles that misalignment. The proposed concrete test, computing FSIM/STSIM between input MRI and Cryosection, would settle whether this concern lands. If those values are already comparable to the reported output-vs-Cryosection values, then the metrics are not measuring colorization quality. I therefore keep the reader's conditional verdict: the paper is promising, but the evaluation must be redone with alignment-aware or misalignment-invariant metrics before the superiority claim can be accepted.","tokens_in":12326,"tokens_out":5006,"duration_ms":50648,"concrete_test":"Compute FSIM and STSIM between the input MRI and its paired Cryosection slice on the reported test set. If these values are already comparable to the reported output-vs-Cryosection values (or low in absolute terms, e.g., below 0.5), the comparison is confounded by residual misregistration. As a second check, re-run Table I using spatially invariant color statistics (e.g., per-segmentation-region color histograms) or after applying the same deformable registration to the colorized output before computing FSIM/STSIM; if the ranking changes, the claimed superiority is an artifact of misalignment.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim (Table I) is not interpretable as stated. In Section IV-A, FSIM and STSIM are described as measuring similarity between the colorized MRI and the Cryosection ground truth, and the paper reports FSIM=0.755 and STSIM=0.901 for the proposed method. However, the Introduction explicitly states that after rigid and B-spline deformable registration, 'Deformations persist in the MRI data ... therefore, the available Cryosection segmentation data cannot be paired with MRI data.' If the modalities cannot be paired, then a colorized MRI that is structurally faithful to the input MRI will be spatially misaligned with the Cryosection reference. Computing pixel-level FSIM/STSIM on misaligned structures measures registration error more than color/texture fidelity. The same confound affects any per-pixel comparison to Cryosection, since the output is aligned to the MRI, not to the Cryosection. The reported numbers therefore do not support the abstract's claim of 'better quantitative results'; they only show similarity to a misaligned reference. This is load-bearing because the quantitative superiority claim rests entirely on these metrics.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a cycle-consistent GAN architecture for colorizing whole-body MRI by transferring colors from Visible Korean Human Cryosection data while preserving the MRI's own structures. The main novelty is a dual-decoder MRI-to-colorized-MRI generator: one decoder builds organ-to-color associations from Cryosection segmentation, and the other performs colorization, with skip-connection compression-activation blocks and a multiscale encoder. The method uses cyclic adversarial loss, an SSIM-based structural loss, and a segmentation loss. The authors report quantitative comparisons on 2,000 test slices against five baselines (ChromaGAN, ColorFormer, DDColor, APS, ALDM), claiming higher SSIM, MS-SSIM, STSIM, FSIM and lower ΔCF, plus qualitative and ablation results.","tokens_in":12572,"tokens_out":3057,"duration_ms":33730,"significance":"If the reported results are reliable, the paper would make a useful contribution to medical image colorization: it addresses a practical scenario of partially registered cross-modal data, avoids requiring MRI segmentation at inference, and proposes a concrete architectural mechanism (dual-decoder fusion with compression-activation skip connections) for organ-consistent color transfer. The ablation study covering cycle consistency, segmentation supervision, pseudo-Cryosection synthesis, and compression-activation is informative. However, two evaluation issues are load-bearing for the central performance claim: the train/test split is described as random slices from a single subject without volume-level separation, which likely leaks nearly duplicate adjacent slices into the test set; and the FSIM/STSIM metrics are computed against Cryosection ground truth that the paper itself says is not pairable with MRI data because deformations persist. These issues mean the quantitative superiority claim in the abstract and Section IV is not currently supported. The architecture is interesting, but the experimental evidence needs substantial revision.","major_comments":[{"comment":"The train/test split is not valid for the claimed generalization. The paper states that 12,000 slices are extracted at random planes from the volumetric MRI, Cryosection, and Segmentation data of a single subject and then split into 10,000 train and 2,000 test samples. Adjacent slices from the same volume are highly correlated, so a random slice split almost certainly places near-duplicate slices in both train and test sets. This can inflate all reported metrics and make the comparison against baselines unreliable. The authors should split by volume or by spatially separated slice blocks, or explicitly show that test slices are not adjacent to training slices. Without this, Table I does not support the superiority claim.","section":"Section IV, dataset construction"},{"comment":"The FSIM and STSIM evaluations are internally inconsistent with the paper's own registration statement. The Introduction states that after rigid and B-spline deformable registration, deformations persist and 'the available Cryosection segmentation data cannot be paired with MRI data.' Yet Section IV-A defines FSIM and STSIM as measuring perceptual and textural similarity 'between ground truth Cryosection and colorized MRI,' and Table I reports FSIM=0.755 and STSIM=0.901 for the proposed method. Because the colorized output is aligned to the input MRI, not to the Cryosection volume, per-pixel FSIM/STSIM scores will largely reflect residual misregistration rather than color/texture fidelity. The reported FSIM/STSIM numbers are therefore not interpretable as evidence of colorization quality. The authors need to either evaluate on registered/test-time-aligned pairs or use metrics that are invariant to the known residual deformation.","section":"Section IV-A and Introduction"},{"comment":"The baseline comparison may be unfair because no tuning details are given. The paper says all five competing methods were retrained on the Visible Korean Human dataset, but it does not report the training budget, hyperparameters, data augmentation, or per-method tuning for ChromaGAN, ColorFormer, DDColor, APS, and ALDM. Differences in training protocol alone can produce large differences in the reported metrics. The authors should provide the same training details for baselines as for their own method, or cite specific configurations and show that each baseline was given a reasonable chance to perform well.","section":"Section IV-B and Table I"},{"comment":"The structural similarity metric is also part of the training objective, which weakens its role as an independent evaluation. Equations (5) and (6) define a multi-scale SSIM loss Lssim that is optimized during training, and Table I then reports SSIM as the headline structural metric. This does not make the comparison meaningless, but it means the SSIM gain over baselines is partly predetermined by the objective. The authors should report a structural metric that is not optimized in the loss, or clearly discuss this overlap in the evaluation. The same concern applies less directly to MS-SSIM, which is SSIM-derived.","section":"Section III-D and Table I"}],"minor_comments":[{"comment":"The sentence 'Our proposed colorization architecture is designed using the generative adversarial network with a cyclic cross-modality adaptation technique [30], as shown in as shown in Fig. 1' contains a duplicated phrase; please fix.","section":"Section III-A"},{"comment":"The notation s, ŝ ∈ R3 → {0,1} does not correctly describe multi-class segmentation maps. The maps should be tensors of shape (height, width, number_of_classes) with values in [0,1] or {0,1}; please revise the notation.","section":"Equation (8)"},{"comment":"The reported ΔCF = -0.001 ± 0.067 for A1 is negative, which is inconsistent with the definition of ΔCF as a difference of colorfulness scores unless negative differences are allowed. Please clarify whether ΔCF is an absolute difference or a signed difference, and ensure the sign conventions are consistent across all rows.","section":"Table II, A1 row"},{"comment":"The standard deviations for ALDM (CF 0.831 ± 0.827, ΔCF 0.086 ± 0.760) are implausibly large and exceed the plausible range of the metrics. Please check the reported values; this may be a typographical error.","section":"Table I, ALDM row"},{"comment":"The caption reads 'Quantitative comparison of MRI colorization with different methods,' but the figure shows qualitative results. The caption should say 'Qualitative comparison.'","section":"Fig. 3 caption"},{"comment":"There is a typo: 'colors in the output should be similar to Crysoction data' should read 'Cryosection data.'","section":"Section IV-A"}],"recommendation":"major_revision","confidential_remarks":"The paper's central architectural idea is worth considering, but the evaluation section needs a major rework before the performance claims can be accepted. The random slice split is a likely leakage problem, and the FSIM/STSIM comparison against misaligned Cryosection ground truth contradicts the authors' own registration caveat. I would recommend the editor ask for a revised version with volume-level evaluation, corrected metrics, and fuller baseline tuning details. If the authors cannot re-run the experiments, the claims should be substantially downgraded."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know about this paper: it has a genuinely new idea — using Cryosection segmentation semantics, delivered through a pseudo-Cryosection decoder, to colorize whole-body MRI without precise registration or MRI segmentation. That is worth reading. But the quantitative claims as reported should not be taken at face value.\n\nWhat's good: the architecture is clearly described, and the ablation study is honest. The authors test five ablations, and the metrics consistently degrade when cycle consistency, segmentation, pseudo-Cryosection, or compression-activation are removed. That suggests each component is doing something. The multiscale module is straightforward. The work addresses a real niche: MRI colorization for anatomy visualization.\n\nNow the soft spots, in order of severity.\n\nFirst, the train/test split. They extract 12,000 slices at random planes from a single subject and split 10,000/2,000. No volume-level separation. Adjacent slices from the same body are near-duplicates. The test set almost certainly contains slices that are essentially the same anatomy as training slices. That inflates every metric and makes the reported superiority over baselines unreliable. This is load-bearing for the central claim.\n\nSecond, the FSIM and STSIM numbers. The paper says in the Introduction that after registration, 'deformations persist ... therefore, the available Cryosection segmentation data cannot be paired with MRI data.' But then Section IV-A computes FSIM and STSIM between the colorized MRI and the ground truth Cryosection. The colorized MRI is aligned to the input MRI, not to the Cryosection. Comparing pixel-level features between misaligned images measures misalignment more than color/texture fidelity. So those numbers do not support the abstract's claim of 'better quantitative results.' The stress-test note is correct, and I don't see a way around it within the paper.\n\nThird, minor: SSIM and MS-SSIM appear in both the training loss and the evaluation metrics, so improvement on them is partly by construction. Also, baseline retuning details are not given, and no code or data is actually available despite a promise. These are fixable.\n\nOverall: the method is plausible and the idea is worth pursuing, but the evaluation needs a major overhaul — proper data separation, registration-aware evaluation, released artifacts. That's a revise-and-resubmit trajectory, not a desk reject. Send it to reviewers who know medical image registration, because they'll catch the FSIM issue immediately.","headline":"Plausible architecture for whole-body MRI colorization, but the reported numbers can't be trusted: the test set likely leaks from training, and FSIM/STSIM compare against a misaligned reference the paper itself says can't be paired.","tokens_in":13064,"tokens_out":2496,"would_cite":false,"duration_ms":22165,"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":"Colorized MRIs can keep their native structure while adopting cryosection colors by routing both outputs through a shared encoder and supervising with segmentation labels.","keywords":["MRI colorization","cross-modal fusion","cycle-consistent GAN","cryosection","structural similarity","squeeze-and-excitation","medical image colorization","multiscale learning"],"falsifier":"Re-run the Table I benchmark with a volume-disjoint split (for example, train on slices from one contiguous body segment and test on non-overlapping segments, or enforce a minimum slice-distance gap between train and test) and recompute the metrics; if the structural and color gains over APS and the other baselines shrink toward zero, the claimed superiority is an artifact of slice-level leakage rather than structural consistency.","tokens_in":12159,"feed_emoji":"🎨","tokens_out":6251,"duration_ms":58005,"temperature":0.7,"pith_summary":"This paper aims to colorize whole-body MRI scans using colors from cryosection anatomy data while preserving the MRI's own structure, without needing precise registration between the two modalities or segmentation of the MRI. The proposed architecture builds the color-structure association through a cycle-consistent GAN whose MRI-to-colorized generator has two decoders: one synthesizes a pseudo-cryosection image from the MRI, and the other uses that association to colorize. A segmentation network trained on cryosection labels supervises the pseudo output, and squeeze-and-excitation-style compression-activation blocks in the skip connections suppress MRI noise and capture organ-level context. The authors report the method outperforms five colorization baselines on structural and color/texture metrics, and they argue this moves medical colorization from partial-body demonstrations to full-body, organ-specific color transfer.","feed_headline":"MRI colorization keeps anatomy intact while borrowing cryosection color","feed_subtitle":"A dual-decoder GAN with segmentation cues outperforms five baselines in structural and color similarity.","key_machinery":"The central object is the dual-decoder colorization generator $G_{m\\to\\hat{c}}$ embedded in a cycle-consistent GAN. One decoder, $F_{x\\to c'}$, synthesizes a pseudo-cryosection image from the shared bottleneck feature $x$, and a fixed U-Net segmenter $S_{c\\to\\hat{s}}$ turns that pseudo image into a label map, so cryosection segmentation semantics can supervise MRI colorization through cross-entropy loss even though the MRI and cryosection volumes are only partially registered. The other decoder, $F_{x\\to\\hat{c}}$, produces the final colorized MRI, and both decoders share the encoder $E_{m\\to x}$; compression-activation blocks on skip connections suppress MRI noise and capture organ-level global information. The total objective is $L = L_{\\text{cyc}} + L_{\\text{ssim}} + L_{\\text{seg}}$.","core_discovery":"The central discovery is that structural consistency in medical colorization can be achieved by weaving cryosection segmentation semantics into a cyclic cross-modality generator without requiring paired, registered data. Concretely, the colorization generator $G_{m\\to\\hat{c}}$ produces both a colorized MRI $\\hat{c}$ and a pseudo-cryosection image $c'$ from the same bottleneck; a fixed U-Net segmentation network maps $c'$ to a label map that is compared with the true cryosection segmentation via cross-entropy, aligning organ regions across modalities. The structural-consistency loss is a multi-patch SSIM between the input MRI and colorized MRI, and channel compression-activation blocks reweight features in skip connections to suppress noise and encode organ-level global context. On the Visible Korean Human data the authors report SSIM 0.863, MS-SSIM 0.866, STSIM 0.901, FSIM 0.755, and $\\triangle$CF 0.005, which they interpret as colorized MRIs that are structurally closest to the input MRI and color/texture closest to the cryosection ground truth.","pith_inferences":["Inference: if the train/test split is not volume-disjoint, the reported margin over baselines may shrink substantially; a nearest-neighbour check between test slices and training slices would reveal how much of the advantage is memorization of near-duplicate anatomy.","Inference: the same dual-decoder-plus-segmentation supervision scheme could transfer to other radiometric-to-anatomical colorization pairs, such as CT to cryosection, or to cross-modality synthesis where one modality has labels and the other does not.","Inference: the compression-activation blocks may also serve as a general denoising mechanism for MRI-to-image translation beyond colorization, since ablation A5 shows their removal erodes fine structure.","Inference: an unsupervised MRI segmentation capability would follow if the pseudo-cryosection decoder were reversed or shared, a direction the authors list as future work."],"forward_implications":["Full-body MRI colorization no longer requires per-organ MRI segmentation or precisely registered MRI-cryosection pairs; the association is learned through a pseudo-cryosynthesis path.","Organ-level color consistency can be enforced by cryosection segmentation labels even when the source MRI has deformations and noise.","The multiscale module lets the same generator colorize inputs at 256x256, 128x128, and 64x64 with limited perceptual loss.","Reported quantitative results (SSIM 0.863, MS-SSIM 0.866, STSIM 0.901, FSIM 0.755, Delta-CF 0.005) position the method above ChromaGAN, ColorFormer, DDColor, APS, and ALDM on this dataset.","Compression-activation in skip connections is what preserves fine structures and organ-specific color variance: removing it (ablation A5) drops SSIM from 0.863 to 0.399 and raises Delta-CF from 0.005 to 0.203."],"supporting_citations":[{"why":"This citation supplies the cycle-consistent adversarial translation backbone that the colorization generator modifies.","marker":"[30]"},{"why":"This citation supplies the compression-activation (squeeze-and-excitation) mechanism used in skip connections for noise suppression and organ-level context.","marker":"[20]"},{"why":"This citation supplies the U-Net segmentation architecture that maps the pseudo-cryosection output to label maps for segmentation loss.","marker":"[31]"},{"why":"This citation supplies the adversarial loss formulation used in the cyclic cross-modality objective.","marker":"[32]"},{"why":"This citation supplies the SSIM measure used both as evaluation metric and as multi-patch structural adaptation loss.","marker":"[33]"},{"why":"This citation supplies the Visible Korean Human cryosection, MRI, and segmentation volumes used for training and evaluation.","marker":"[26]–[28]"}],"fun_headline_variants":["MRI colorization with cryosection color, structure preserved via fusion","No registration needed: MRI gets cryosection color while structure stays intact","Fusing cryosection segmentation semantics for structure-preserving MRI colorization","Cross-modal fusion keeps MRI anatomy while adding cryosection color"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the 2,000 test slices are independent of the 10,000 training slices, even though all 12,000 slices come from random planes within a single subject's registered volumes, so adjacent slices in the test set may be near-duplicates of training slices.","fun_headline_variants_meta":{"raw":{"variants":["MRI colorization with cryosection color, structure preserved via fusion","No registration needed: MRI gets cryosection color while structure stays intact","Fusing cryosection segmentation semantics for structure-preserving MRI colorization","Cross-modal fusion keeps MRI anatomy while adding cryosection color"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000584,"raw_usage":{"total_tokens":2737,"prompt_tokens":928,"completion_tokens":1809,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":1742}},"tokens_in":544,"tokens_out":1809,"duration_ms":12826,"temperature":1.0,"reasoning_tokens":1742,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:21:07.437568+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Table I benchmark with a volume-disjoint split (for example, train on slices from one contiguous body segment and test on non-overlapping segments, or enforce a minimum slice-distance gap between train and test) and recompute the metrics; if the structural and color gains over APS and the other baselines shrink toward zero, the claimed superiority is an artifact of slice-level leakage rather than structural consistency.","supporting_citations":[{"cited_title":"Squeeze-and-Excitation networks,","cited_arxiv_id":null,"evidence_quote":"This citation supplies the compression-activation (squeeze-and-excitation) mechanism used in skip connections for noise suppression and organ-level context."},{"cited_title":"U-Net: Convolutional net- works for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"This citation supplies the U-Net segmentation architecture that maps the pseudo-cryosection output to label maps for segmentation loss."}],"review_version":1}