{"id":"a49e4dd2-04db-4212-b9cb-ecadde4fb622","arxiv_id":"2607.07553","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":7,"one_line_summary":"AA-ViT synthesizes contrast-enhanced brain MRI from pre-contrast T1/T2/FLAIR using edge-aware dense encoding and frequency-guided losses, achieving higher PSNR/SSIM than four baselines on BraTS2021.","lead":"The paper proposes AA-ViT, a vision transformer that synthesizes contrast-enhanced brain MRI from non-contrast scans using edge-aware encoders and frequency-domain losses. A smart generalist might read it because eliminating gadolinium contrast agents could reduce MRI costs and patient risks.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"No inter-method significance testing in Table 1; with SD ≈ 4.9 dB, the 0.87 dB PSNR improvement over ResViT may not be statistically significant, undermining the core quantitative claim.","rationale":"The reader's verdict of CONDITIONAL is appropriate, but for a partially different reason. The reader focused on the clinical translation gap (19 cases, no non-inferiority margin, no blinded reader study against real contrast). This is a valid concern, but the authors themselves acknowledge it explicitly in the conclusion, and the paper frames the clinical evaluation as 'preliminary.' A more load-bearing concern is that the core quantitative claim — outperforming SOTA on PSNR/SSIM — is not backed by inter-method significance testing, and the standard deviations are large enough (~4.9 dB) that the 0.87 dB improvement over ResViT could be non-significant. The ablation study demonstrates that each component contributes significantly over the baseline, which is encouraging, but this does not substitute for testing against competing methods. The loss function also has a potential redundancy: L_L1 (Eq. 4, λ1 term) and L_err (Eq. 5) both appear to be MAE, which is confusing but not necessarily incorrect if different implementations are used. Loss weight values are not reported, reducing reproducibility. Additionally, BraTS2021 images are skull-stripped, co-registered, and resampled to 1mm isotropic, making the synthesis task substantially easier than real clinical CEMRI — a limitation the authors acknowledge but which bears on the generalizability of both the quantitative and clinical results. Overall, the paper presents a reasonable architectural contribution with a well-structured ablation, but the headline SOTA claim needs statistical validation. CONDITIONAL remains the correct verdict: the ML contribution is plausible but not fully established, and the clinical claim is preliminary as the reader noted.","tokens_in":7884,"tokens_out":2390,"duration_ms":117095,"concrete_test":"Run a paired Wilcoxon signed-rank test (or paired t-test) on per-case PSNR and SSIM values between AA-ViT and each baseline (ResViT, I2I-Mamba, TSF-Seq2Seq, MU-Diff) on the 328-subject test set. Report p-values and effect sizes (e.g., Cohen's d). If the improvement over ResViT is not significant at p < 0.05, the headline claim of outperforming SOTA weakens substantially. Also report the percentage of test cases where AA-ViT outperforms each baseline to assess consistency.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that AA-ViT outperforms four SOTA methods on PSNR and SSIM. Table 1 reports AA-ViT PSNR = 27.790 ± 4.927 vs. ResViT 26.919 ± 4.657, a ~0.87 dB gap. The standard deviations are enormous relative to this difference. Critically, while the ablation study (Table 2) uses Wilcoxon signed-rank tests with p-values, the main comparison table (Table 1) reports no significance tests at all. Without a paired test on per-case results, it is unclear whether the improvement over ResViT (or I2I-Mamba at 26.615) is statistically meaningful or within noise. The large SDs suggest high inter-case variance — some cases likely show large gains while others may show regressions, and the mean difference could be driven by a subset of cases. This is the load-bearing issue because the entire paper's contribution rests on the claim of superior quantitative performance; if the improvement is not statistically significant, the contribution narrows to the architectural components and ablation analysis alone. The reader correctly identified the clinical evaluation gap but did not flag this more fundamental issue with the quantitative comparison.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"This paper proposes AA-ViT, an anatomically aware vision transformer for synthesizing contrast-enhanced brain MRI (T1ce) from pre-contrast modalities (T1, T2, FLAIR). The method introduces a Residual Dense Edge Block (RDEB) in the encoder for edge-aware feature extraction and a multi-component loss combining L1, adversarial, error-map, edge (Sobel), and frequency-domain (FFT) terms. The method is evaluated on BraTS2021 against four baselines (ResViT, I2I-Mamba, TSF-Seq2Seq, MU-Diff) using PSNR and SSIM, with an ablation study using Wilcoxon signed-rank tests, a CKA representation analysis, and a preliminary clinical evaluation by four clinicians on 19 cases.","tokens_in":8050,"tokens_out":1870,"duration_ms":152332,"significance":"The paper addresses a clinically relevant problem (gadolinium-free CEMRI synthesis) and includes a preliminary clinical reader study, which is uncommon in this literature. The ablation study is properly designed with statistical significance testing (Wilcoxon, p<0.001), and the CKA analysis provides representation-level evidence beyond pixel metrics. The architectural components (RDEB, edge/FFT losses) are clearly motivated and individually validated. However, the main comparison against baselines lacks the same statistical rigor applied in the ablation, which weakens the central outperformance claim.","major_comments":[{"comment":"Table 1 (main comparison): The central claim that AA-ViT outperforms SOTA methods rests on mean±SD values with no inter-method significance testing. The PSNR gap over ResViT is 0.87 dB (27.790±4.927 vs. 26.919±4.657) with standard deviations of ~4.7–4.9 dB. Given this ratio of effect size to variance, a paired statistical test on per-case results is essential to determine whether the improvement is meaningful or within noise. Notably, Table 2 (ablation) already employs Wilcoxon signed-rank tests, so the infrastructure exists. Without paired tests in Table 1, the core quantitative claim is unsupported. This is the most load-bearing issue in the paper.","section":null},{"comment":"Section 2.2, Eq. (4)–(8): The loss function introduces seven free parameters (λ1–λ5, α, r_frac) described as 'empirically chosen.' No sensitivity analysis is provided for any of these. Since the edge influence scalar α and the high-frequency threshold r_frac directly govern the anatomical and frequency guidance that distinguish AA-ViT from the ResViT baseline, the absence of any sensitivity study makes it difficult to assess robustness. At minimum, a sensitivity analysis for α and r_frac should be included, as these are the most architecturally novel hyperparameters.","section":null},{"comment":"Section 3.2 (Clinical Evaluation): The clinical evaluation on 19 cases with a mean Likert score of 3.94/5 is presented as 'preliminary clinical validation.' While the authors acknowledge limitations in the conclusion, the study design has specific gaps that should be addressed or more clearly bounded: (1) no power analysis or non-inferiority margin is defined; (2) there is no comparison against real contrast-enhanced images (i.e., readers were not asked to distinguish synthetic from real, nor rate real images for comparison); (3) it is unclear whether the evaluation was blinded to image origin. The claim of 'clinical validation' should be scaled back to 'preliminary clinical assessment' unless a blinded comparison with real images is included.","section":null}],"minor_comments":[{"comment":"Abstract and Section 3.1: The abstract states 'three neuroradiologists and a neurosurgeon' while the clinical evaluation section also says 'three neuroradiologists and one neurosurgeon' — consistent, but the number of readers (n=4) and cases (n=19) should be stated in the abstract for proportionality.","section":null},{"comment":"Table 1: PSNR is reported as 27.790 in the table but 27.71 in the text (Section 3.1, 'Quantitative Results'). Similarly SSIM is 0.930 in the table but 0.929 in the text. Please reconcile.","section":null},{"comment":"Section 2.1, Eq. (3): The notation C_{1×1} is introduced without explicit definition as a 1×1 convolution; this should be clarified for consistency with C_l in Eq. (1).","section":null},{"comment":"Section 2.2, Eq. (6): The notation 'C_l' is used for convolution with Sobel kernels, but C_l was previously defined in Eq. (1) as a 3×3 convolution in the dense block. Using the same symbol for different operations is confusing; consider distinct notation.","section":null},{"comment":"Figure 2: The CKA values (0.948±0.102 for AA-ViT vs. 0.486±0.311 for ResViT) are mentioned in the text but the figure caption does not specify which layers correspond to which rows/columns. Adding axis labels or layer indices would improve interpretability.","section":null},{"comment":"Section 3.1: The dataset preprocessing mentions 'axial slices were scored by intensity across non-zero voxels, retaining the top half (≤100 per subject).' This selection criterion could introduce bias toward high-intensity (enhancing) slices; a brief justification would help.","section":null},{"comment":"Reference [3] and [13] both appear to be 2025 MICCAI papers; please verify these are published or update to the correct venue/status.","section":null}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about missing significance tests in Table 1 is well-founded and is the primary reason for the major revision recommendation. The authors clearly have per-case results (they ran paired Wilcoxon tests for the ablation), so this should be straightforward to address. The clinical evaluation, while preliminary, is a genuine effort that adds value; I would not want the authors to remove it, but the framing should be more conservative. The paper is within scope for a medical imaging venue."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive review. The referee raises three major points: (1) the need for paired inter-method significance testing in Table 1, (2) the absence of a sensitivity analysis for the edge influence scalar α and the high-frequency threshold r_frac, and (3) concerns about the scope and design of the clinical evaluation, particularly the use of the phrase 'clinical validation.' We agree with all three points and will revise the manuscript accordingly. Specifically, we will add Wilcoxon signed-rank tests to the main comparison table, include a sensitivity analysis for α and r_frac, and reframe the clinical evaluation as a 'preliminary clinical assessment' with clearer bounding of its limitations. We note one honest constraint: the clinical evaluation was conducted as a single-arm reader study (synthetic images only) and was not designed with blinding to image origin or a real-image comparator; we cannot retroactively add these design elements to the existing data, so we will instead clearly state these as limitations and scope the claims accordingly.","responses":[{"response":"The referee is correct. We already apply Wilcoxon signed-rank tests in the ablation table (Table 2), so the infrastructure and per-case results are available. We will add the same paired Wilcoxon signed-rank tests to Table 1 for all pairwise comparisons between AA-ViT and each baseline (ResViT, I2I-Mamba, TSF-Seq2Seq, MU-Diff) on both PSNR and SSIM. We will report p-values (or significance asterisks) directly in the table. If any pairwise comparison does not reach significance after correction for multiple comparisons, we will state this transparently rather than claiming outperformance on that metric. We agree that without these tests the central claim is not adequately supported, and we will temper the language in the abstract and results section accordingly if any comparison fails to reach significance.","revision_made":"yes","referee_comment":"Table 1 (main comparison): The central claim that AA-ViT outperforms SOTA methods rests on mean±SD values with no inter-method significance testing. The PSNR gap over ResViT is 0.87 dB (27.790±4.927 vs. 26.919±4.657) with standard deviations of ~4.7–4.9 dB. Given this ratio of effect size to variance, a paired statistical test on per-case results is essential to determine whether the improvement is meaningful or within noise. Notably, Table 2 (ablation) already employs Wilcoxon signed-rank tests, so the infrastructure exists. Without paired tests in Table 1, the core quantitative claim is unsupported. This is the most load-bearing issue in the paper."},{"response":"We agree. We will add a sensitivity analysis for α and r_frac, varying each across a reasonable range (e.g., α ∈ {0.1, 0.25, 0.5, 0.75, 1.0} and r_frac ∈ {0.1, 0.2, 0.3, 0.4, 0.5}) and reporting PSNR and SSIM for each setting. This will be included as a new table or figure in the revised manuscript. We will also briefly note the rationale for the chosen default values based on this analysis. We concur that α and r_frac are the most architecturally novel hyperparameters and that demonstrating robustness (or documenting sensitivity) is important for the reader to assess the method's practical reliability.","revision_made":"yes","referee_comment":"Section 2.2, Eq. (4)–(8): The loss function introduces seven free parameters (λ1–λ5, α, r_frac) described as 'empirically chosen.' No sensitivity analysis is provided for any of these. Since the edge influence scalar α and the high-frequency threshold r_frac directly govern the anatomical and frequency guidance that distinguish AA-ViT from the ResViT baseline, the absence of any sensitivity study makes it difficult to assess robustness. At minimum, a sensitivity analysis for α and r_frac should be included, as these are the most architecturally novel hyperparameters."},{"response":"We agree with the referee's assessment. We will make the following changes: (1) Replace all instances of 'clinical validation' with 'preliminary clinical assessment' throughout the manuscript, including the abstract, Section 3.2, and the conclusion. (2) Explicitly state in Section 3.2 that the evaluation was a single-arm reader study on synthetic images only, with no real-image comparator and no formal blinding to image origin. (3) Add a clear statement that no power analysis or non-inferiority margin was pre-specified. (4) Frame the 3.94/5 Likert score as exploratory evidence of perceived image quality and diagnostic utility, not as validation of clinical equivalence or non-inferiority to real contrast-enhanced MRI. We acknowledge that a properly powered, blinded, real-versus-synthetic comparison would be needed to support stronger claims, and we will list this as a priority for future work. We cannot retroactively add blinding or a real-image comparator to the existing study data, so we will bound the claims to what the study design supports.","revision_made":"yes","referee_comment":"Section 3.2 (Clinical Evaluation): The clinical evaluation on 19 cases with a mean Likert score of 3.94/5 is presented as 'preliminary clinical validation.' While the authors acknowledge limitations in the conclusion, the study design has specific gaps that should be addressed or more clearly bounded: (1) no power analysis or non-inferiority margin is defined; (2) there is no comparison against real contrast-enhanced images (i.e., readers were not asked to distinguish synthetic from real, nor rate real images for comparison); (3) it is unclear whether the evaluation was blinded to image origin. The claim of 'clinical validation' should be scaled back to 'preliminary clinical assessment' unless a blinded comparison with real images is included."}],"tokens_in":7811,"tokens_out":1269,"duration_ms":151236,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"Two things matter here. First, the architectural contribution — a Sobel-edge-aware dense encoder block plus error-map, edge, and FFT losses bolted onto ResViT — is legitimate but incremental. Each piece is well-established; the combination is sensible and the ablation shows each component helps with statistical significance (Wilcoxon, p<0.001). Second, the clinical evaluation (19 cases, 4 clinicians, Likert scale) is too small to support the paper's framing around gadolinium replacement, and the authors acknowledge this honestly in their limitations section. So the ML claim is modestly supported; the clinical claim is aspirational. The stress-test note flags a real problem: Table 1 reports no inter-method significance tests. With PSNR SDs around 4.9 dB and a 0.87 dB gap over ResViT, a paired test on per-case results is essential. The ablation table uses Wilcoxon tests, so the authors clearly know how to do this — its absence in the main comparison is a gap, not an oversight they'd defend. That said, the improvement is consistent across both metrics and the ablation is well-structured. The CKA analysis (Kornblith et al.) showing higher representation similarity for AA-ViT vs. ResViT is a nice addition — it gives representation-level evidence beyond pixel metrics, and the final decoder layer CKA jump (0.948 vs. 0.486) is notable. Loss weights are not reported, which hurts reproducibility. The paper is for researchers working on medical image synthesis and multi-modal MRI translation. It deserves a serious referee who can ask for the missing significance tests and push back on clinical framing. I'd accept for peer review with a request to add paired statistical tests to Table 1 and tone down the clinical claims to match the evidence.","headline":"Solid incremental architecture for contrast-enhanced MRI synthesis; main comparison table lacks significance testing and clinical evaluation is preliminary","tokens_in":8944,"tokens_out":444,"would_cite":false,"duration_ms":141451,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Synthetic Brain Scans Without Contrast Dye Edge Closer to Clinical Use","keywords":["contrast-enhanced MRI synthesis","vision transformer","anatomical edge awareness","frequency-domain loss","gadolinium-free imaging","brain tumor MRI","BraTS 2021","generative medical imaging"],"falsifier":"If a focused lesion-level evaluation found that AA-ViT systematically hallucinated or missed contrast enhancement in clinically critical regions (e.g., active tumor margins) at rates that would change clinical decisions, the core claim of anatomically faithful synthesis would be undermined regardless of global PSNR/SSIM improvements.","tokens_in":7938,"feed_emoji":"🧠","tokens_out":1529,"duration_ms":147398,"temperature":0.7,"pith_summary":"This paper argues that explicitly forcing a neural network to care about anatomical edges and high-frequency detail — rather than letting it learn those features implicitly — produces more faithful synthetic contrast-enhanced brain MRI from pre-contrast scans. The authors build on an existing transformer-based generator (ResViT) and add two mechanisms: a Residual Dense Edge Block that extracts gradient-based edge maps using Sobel kernels and fuses them into the encoder's feature representations, and a multi-component training objective that penalizes errors in edge alignment, pixel-wise reconstruction, and frequency-domain fidelity via a high-pass Fourier mask. The combined effect is that the generator is steered toward preserving tumor boundaries and fine anatomical structures that prior methods blur or hallucinate. On the BraTS 2021 dataset (1,251 subjects), the resulting model, AA-ViT, achieves PSNR of 27.79 and SSIM of 0.930, outperforming four state-of-the-art methods spanning transformer, state-space, and diffusion architectures. A preliminary clinical evaluation by three neuroradiologists and one neurosurgeon across 19 glioma cases yielded a mean Likert score of 3.94 out of 5 across image quality, diagnostic utility, lesion visibility, clinical confidence, and added informational value. The authors position this as a step toward eliminating gadolinium-based contrast agents, which carry risks of brain deposition and nephrogenic systemic fibrosis and are contraindicated in patients with renal impairment.","feed_headline":"Edge-Aware AI Synthesizes Contrast Brain MRI Without Dye","feed_subtitle":"Explicit anatomical edge and frequency guidance lets a transformer generate gadolinium-free scans that beat four SOTA methods on pixel and初步","key_machinery":"Residual Dense Edge Block (RDEB): an encoder module that extracts anatomical edge maps via Sobel kernels and fuses them with densely connected convolutional features through a learnable projection, so that structural boundary information is explicitly carried into the generator's representations rather than left to be discovered implicitly.","core_discovery":"The central discovery is that explicit anatomical edge supervision and frequency-domain loss, when combined with a dense edge-aware encoder, yield measurably better synthesis of contrast-enhanced T1 MRI from pre-contrast T1, T2, and FLAIR inputs than architectures that rely on implicit feature learning. The Residual Dense Edge Block extracts structural edge information via Sobel convolution and injects it into the dense feature pathway, while the anatomically aware loss simultaneously penalizes pixel error, adversarial realism, edge misalignment, and high-frequency spectral deviation. An ablation study shows each component contributes statistically significant improvement (Wilcoxon signed-r,","pith_inferences":["The Sobel-based edge extraction is a fixed, hand-crafted prior; a natural extension would be to learn task-specific edge detectors or to incorporate multi-scale structural priors (e.g., from anatomical segmentation networks) that could capture boundaries Sobel misses in low-contrast regions.","The frequency-domain loss uses a single radial high-pass threshold; adaptive or learned frequency masks that target clinically relevant frequency bands (e.g., those corresponding to lesion texture scales) could improve diagnostic fidelity further.","The 19-case clinical evaluation, while a step beyond pure metric comparison, is too small to detect rare but dangerous failure modes such as hallucinated enhancement in critical regions — a non-inferiority study with lesion-level ground truth would be needed before any clinical deployment claim.","The slice-wise 2D processing discards inter-slice volumetric context; extending RDEB and the frequency loss to 3D would likely improve consistency across slices, which is clinically important for volumetric tumor assessment."],"forward_implications":["If the edge-and-frequency-guidance approach generalizes beyond BraTS 2021, it could become a standard ingredient in medical image synthesis pipelines where anatomical boundary fidelity is clinically critical — not just for brain MRI but for any contrast-enhanced modality.","The CKA analysis showing AA-ViT's final decoder layer reaching 0.948 representation similarity to real CEMRI (vs. 0.486 for ResViT) suggests that explicit edge priors shift learned representations closer to the ground-truth distribution at a deep architectural level, not just at the pixel output.","The preliminary clinician ratings (3.94/5) indicate that synthetic images are approaching but have not reached the threshold where clinicians would confidently use them as replacements for real contrast scans — the gap between 'useful adjunct' and 'safe substitute' remains the key clinical barrier.","The authors' own acknowledgment that current metrics do not assess missed or hallucinated enhancement points to a need for lesion-level evaluation protocols specific to synthetic contrast MRI, since global similarity metrics can mask clinically dangerous failures."],"fun_headline_variants":["Anatomy-guided transformer creates contrast brain MRIs without dye","Edge and frequency guidance improve dye-free brain MRI synthesis","Anatomical edge loss sharpens synthetic contrast brain MRI","Explicit edge awareness improves dye-free brain MRI synthesis","Structure-guided AI synthesizes gadolinium-free brain MRI"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The paper assumes that higher pixel-similarity scores (PSNR, SSIM) and favorable clinician ratings on 19 cases are a meaningful proxy for clinical safety, when in fact the authors themselves note that current metrics do not assess whether the model misses or fabricates contrast enhancement in tumor regions — exactly the errors that would matter most in practice.","fun_headline_variants_meta":{"raw":{"variants":["Anatomy-guided transformer creates contrast brain MRIs without dye","Edge and frequency guidance improve dye-free brain MRI synthesis","Anatomical edge loss sharpens synthetic contrast brain MRI","Explicit edge awareness improves dye-free brain MRI synthesis","Structure-guided AI synthesizes gadolinium-free brain MRI","Frequency and edge supervision yield better synthetic brain MRIs"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":1078,"prompt_tokens":629,"completion_tokens":449,"prompt_tokens_details":null},"tokens_in":629,"tokens_out":449,"duration_ms":18253,"temperature":1.0,"reasoning_tokens":416,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T07:03:00.722139+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If a focused lesion-level evaluation found that AA-ViT systematically hallucinated or missed contrast enhancement in clinically critical regions (e.g., active tumor margins) at rates that would change clinical decisions, the core claim of anatomically faithful synthesis would be undermined regardless of global PSNR/SSIM improvements.","supporting_citations":[],"review_version":1}