{"id":"56e11001-8632-480f-92f5-229a79213f79","arxiv_id":"2507.21349","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A deep learning framework using deep registration and a transformer enhancer improves prior-informed brain MRI reconstruction quality and speed.","lead":"This paper describes a deep learning pipeline that reconstructs faster brain MRI scans by using a patient's earlier scan as a guide, combining a neural reconstruction model, a fast registration step, and a transformer-based enhancer. It reports better image quality and shorter registration time than an earlier approach, making prior-informed MRI reconstruction more practical for clinical use.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed p<0.05 superiority over the Souza et al. baseline is unverified because the paper does not show the baseline was re-run on the same 18-subject test set; a paired Wilcoxon test requires shared paired measurements.","rationale":"The reader's weakest assumption identifies exactly the most load-bearing risk: the comparison with the prior Souza et al. baseline may not be matched. My reading of the manuscript confirms this concern and sharpens it. The paper describes the proposed training procedure in detail but is silent on whether the 'Enhanced [Souza et al., UNet]' baseline was retrained under the same conditions or whether its quantitative results were taken from the earlier publication. Because the central claim is a paired statistical comparison (Wilcoxon signed-rank test), the validity of that test depends on having per-subject paired measurements on the same test set. Without that, the p < 0.05 claim cannot be interpreted. The presence of the 'Enhanced [Ours, UNet]' ablation partially mitigates the concern, since it shows the transformer component helps under the proposed pipeline, but it does not establish superiority over the prior method as claimed. The manuscript also has other inconsistencies, such as the expert-rating summary contradicting detailed ratings and dataset slice-count ambiguities, but these are secondary to the baseline-matching question. The code is publicly available, which makes the concrete test feasible and is a point in the paper's favor. I do not see evidence of intentional misrepresentation; the issue is missing specification and potentially missing re-evaluation. Therefore, the verdict should remain conditional on the authors clarifying or redoing the baseline comparison, not a rejection of the underlying idea.","tokens_in":12445,"tokens_out":6599,"duration_ms":75489,"concrete_test":"Inspect the public GitHub repository (github.com/amirshamaei/longitudinal-mri-deep-recon) for evaluation scripts and data-split files. Determine whether 'Enhanced [Souza et al., UNet]' results are produced by running the original WW-net + FSL FLIRT + UNet pipeline on the same 18 test subjects used for the proposed methods, or whether the numbers are imported from the original publication. If they are imported, re-run that baseline on the same test set (or obtain per-subject predictions from the authors) and recompute the paired Wilcoxon signed-rank test for SSIM, PSNR, and NMSE at all four acceleration factors. If the baseline was re-run, verify that the identical train/validation/test split, preprocessing, early-stopping criterion, and evaluation code were used. Also record the CPU/GPU model and input image dimensions for both the 4.1 s and the 90 s registration timings.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the proposed E2E-VarNet + EasyReg + transformer framework outperforms the previously proposed linearly-registered UNet-enhanced reconstruction rests on a paired statistical comparison against 'Enhanced [Souza et al., UNet]' (Section 3, Figure 2, p < 0.05, Wilcoxon signed-rank test). The paper never states that this baseline was re-trained or re-evaluated on the same 18-subject test set under identical preprocessing, train/validation/test split, and hardware. Section 2.4 describes training and validation splits for the proposed models, but for the Souza et al. baseline it only says the experiments were 'designed to compare' with that method and that the same dataset was used. If the baseline numbers are taken from the original publication, which used its own data split, then a paired signed-rank test is not well-defined: there are no paired per-subject measurements on a common test set. The comparison also conflates three independent changes at once: initial reconstruction (WW-net IKIK vs E2E-VarNet), registration (FSL FLIRT vs EasyReg), and enhancement network (UNet vs transformer). The paper's own controlled ablation 'Enhanced [Ours, UNet]' isolates the transformer contribution, but that is not the headline comparison. Therefore, the statistically significant quality advantage over the prior method is currently unverified. The secondary speed claim (4.1 s vs 90 s registration time) also lacks specification of hardware (GPU vs CPU), image size, and whether timing was measured on the same machine; this does not undermine the quality claim but should be clarified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper extends prior-informed MRI reconstruction by replacing linear registration (FSL FLIRT) with a deep registration model (EasyReg) and the UNet enhancement network with a transformer-based enhancement network, using E2E-VarNet as the initial reconstruction. The authors validate on a longitudinal T1-weighted dataset with 18 test subjects at acceleration factors R5-R20, comparing the proposed \"Enhanced [Ours, Transformers]\" against non-enhanced reconstructions, their own UNet-enhanced variant, and the previously proposed \"Enhanced [Souza et al., UNet]\". They report higher SSIM/PSNR and lower NMSE, improved downstream SynthSeg segmentation Dice and volumetric agreement, expert reader quality ratings, and a registration time reduction from about 90 s (FSL FLIRT) to 4.1 s (EasyReg). The code is publicly available. The central scientific claim is that the proposed framework is superior to prior prior-informed reconstruction methods, with statistical significance assessed via the Wilcoxon signed-rank test.","tokens_in":12744,"tokens_out":5264,"duration_ms":60019,"significance":"If the claims hold, this is a useful incremental contribution: it shows that replacing linear registration with a learning-based registration and a UNet with a transformer in a prior-informed MRI reconstruction pipeline improves reconstruction quality, reduces registration time, and preserves downstream segmentation accuracy. Strengths include a held-out test set, multiple quantitative metrics, a downstream segmentation analysis, a controlled ablation of the transformer versus UNet enhancement (\"Enhanced [Ours, UNet]\"), and public code. The use of the prior scan as an input, not as a fitted constant, avoids circularity in the evaluation. The main vulnerabilities are the potentially unmatched external baseline comparison and the conflation of three architectural changes in the headline comparison; these are fixable with additional experiments. The novelty is moderate but appropriate for a specialized MRI methods journal.","major_comments":[{"comment":"The headline claim that the proposed method outperforms \"Enhanced [Souza et al., UNet]\" (p < 0.05, Wilcoxon signed-rank test) is not supported as stated, because the manuscript does not establish that this baseline was re-trained and evaluated on the same 18-subject test set under identical train/validation/test splits and preprocessing. Section 2.4 describes training and validation splits for the proposed models, but for the Souza et al. baseline it only says the experiments were \"designed to compare\" and that the same dataset was used. If the baseline metrics are taken from Ref. [12], which used a different data split, then a paired signed-rank test has no defined pairs. Please either rerun the Souza et al. pipeline under the same protocol or explicitly state and justify the source of the baseline numbers; without this, the central superiority claim is unverified.","section":"Section 2.4, Figure 2"},{"comment":"The comparison between \"Enhanced [Ours, Transformers]\" and \"Enhanced [Souza et al., UNet]\" simultaneously changes the initial reconstruction network (WW-net IKIK vs. E2E-VarNet), the registration method (FSL FLIRT vs. EasyReg), and the enhancement network (UNet vs. transformer). The controlled ablation \"Enhanced [Ours, UNet]\" isolates the transformer contribution but does not isolate the registration contribution. Consequently, the paper's claim that the deep registration provides improved accuracy is not demonstrated; only its speed advantage is shown. Please add an ablation that varies only the registration component (e.g., E2E-VarNet + FLIRT + transformer) to attribute the reconstruction gains appropriately.","section":"Section 3 (Table 1, Figure 2)"},{"comment":"The processing-time comparison between EasyReg (4.1 s) and FSL FLIRT (90 s), and the reconstruction times in Table 1, do not specify the hardware (GPU model vs. CPU), image dimensions, or whether times include preprocessing, model loading, or data transfer. Without these details, the speed advantage and the claim of suitability for real-time clinical applications are not reproducible. Please state the exact evaluation environment and define what each reported time includes.","section":"Section 3, Table 1"},{"comment":"The test set consists of 18 subjects, so subject-level statistical power is limited, yet only aggregate means and a blanket \"p < 0.05\" are reported; no per-subject error bars, exact p-values, or effect sizes are given. Please report the distribution of per-subject metrics (e.g., error bars or box plots) and exact p-values for each metric and acceleration factor, and state whether any multiple-comparison correction was applied.","section":"Section 2.4, Section 3, Figure 2"}],"minor_comments":[{"comment":"Notation is inconsistent: Eq. (4) defines \"PSreg\" but Eq. (5) uses \"Preg\" for the same quantity; please use one symbol throughout.","section":"Section 2.3"},{"comment":"The metric is introduced as NRMSE in Section 2.4 but labeled NMSE in Figure 2 and the surrounding text; please standardize the abbreviation.","section":"Section 2.4, Figure 2"},{"comment":"In the ablation study description, \"evaluating the generalizability of your method\" should read \"our method\".","section":"Section 2.4"},{"comment":"Reference [19] is a diffusion transformer paper; the architectural description of the enhancement network should cite the vision transformer reference [14] consistently rather than the diffusion paper as the source of the ViT-based design.","section":"Section 2.2, References [14], [19]"},{"comment":"No inter-rater agreement (e.g., Cohen's kappa) is reported for the two expert readers; given only six subjects, such a measure would help interpret the reliability of the ratings.","section":"Section 3.1"},{"comment":"The sentence \"These findings underscore the effectiveness of ... (Table 1)\" cites Table 1, but Table 1 lists model components and processing times; the quantitative reconstruction metrics appear in Figure 2. Please correct the reference.","section":"Section 4, first paragraph"}],"recommendation":"major_revision","confidential_remarks":"The controlled comparisons within the paper (Transformers vs. UNet enhancement, both with E2E-VarNet and EasyReg) appear internally consistent, and the non-enhanced comparisons are useful. The main risk is the external baseline: if the Souza et al. numbers were not obtained on the same test set, the headline p-value claim is invalid. I would ask the authors to either rerun that baseline or explicitly reframe the comparison as \"relative to reported values.\" The scope is a good fit for Magnetic Resonance Imaging."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a workmanlike extension of Souza et al.'s prior-informed MRI reconstruction. The new bit is swapping linear registration (FSL FLIRT) for deep registration (EasyReg) and swapping the U-Net enhancer for a transformer, plus switching the initial reconstruction to E2E-VarNet. The advertised 90 s to 4.1 s registration time is a concrete, practically useful gain if it holds on comparable hardware.\n\nWhat the paper does well: it includes an ablation that isolates the transformer contribution (Enhanced [Ours, UNet] vs Enhanced [Ours, Transformers]), a second ablation replacing subject-specific priors with an atlas, downstream segmentation analysis with Dice and volumes, and a public code release. The limitations section is honest, especially about the small healthy-subject dataset and no new-pathology evaluation. Self-citation is transparent.\n\nThe soft spot is the headline comparison. The paper claims p<0.05 superiority over 'Enhanced [Souza et al., UNet]' using a paired Wilcoxon test, but it never states that the Souza baseline was re-trained or re-evaluated on the same 18-subject test set with the same preprocessing and split. If those numbers come from the original paper, a paired test is not well-defined. The design also changes three things at once: initial reconstruction, registration, and enhancer. The controlled ablation handles the enhancer, but not the full comparison. This doesn't kill the paper—the speed-up stands on its own, and the transformer-vs-UNet ablation is meaningful—but the 'superior to existing methods' claim needs re-running the baseline on the same split before I'd trust it.\n\nSmaller issues: no error bars in the text (figures may have them, but the text doesn't say), expert-rating prose overstates 'consistently' when the detailed numbers show the transformer at R5 scored below non-accelerated on fine anatomical detail, and the subject-count accounting for the enhancement network overlaps with the E2E-VarNet training set—probably harmless, but the split description is muddled.\n\nBottom line: worth engaging, worth a serious referee. I'd ask for a re-run of the Souza baseline and paired error bars before acceptance. The registration speed-up alone justifies publication somewhere, but the quality comparison should be made defensible.","headline":"Useful incremental extension of prior-informed MRI reconstruction with a real registration speed-up, but the headline superiority claim needs the baseline re-run on the same test set.","tokens_in":13301,"tokens_out":3344,"would_cite":true,"duration_ms":34920,"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":"This paper argues that prior subject-specific MRI scans improve accelerated reconstruction when the prior is aligned by a fast deep-learning registration model and refined by a transformer-based enhancer, cutting registration time from…","keywords":["MRI reconstruction","deep learning","prior-informed reconstruction","image registration","transformer","brain MRI","longitudinal imaging","accelerated MRI"],"falsifier":"Retrain the previous linearly registered UNet-enhanced method on the exact same training, validation, and test split, preprocessing, and hardware used here, then compare SSIM, PSNR, and NMSE on the same test subjects; if the proposed method no longer wins the Wilcoxon signed-rank test at p < 0.05, the central claim fails. A second check is whether expert readers can still detect the improved reconstructions as anatomically faithful in patients with new lesions.","tokens_in":12278,"feed_emoji":"🧠","tokens_out":6353,"duration_ms":70710,"temperature":0.7,"pith_summary":"MRI scans are slow, and reconstructing a fast, undersampled scan is harder. This paper argues that a patient's previous MRI scan can be used to help, provided the old scan is aligned to the new one quickly and the fusion is done with a transformer-based network. The authors build a pipeline with an initial reconstruction network, a deep-learning registration step, and a transformer enhancer, then test it on 2,808 T1-weighted brain images from 18 subjects at acceleration factors R5, R10, R15, and R20. They report statistically significant gains in SSIM, PSNR, and NMSE over both non-enhanced reconstructions and a previously proposed linearly registered UNet-enhanced baseline, with the gap growing at higher acceleration. Registration time drops from about 90 seconds to 4.1 seconds per subject, and downstream brain segmentation also improves.","feed_headline":"Transformer-enhanced MRI reconstruction beats prior method at R5-R20","feed_subtitle":"Fast deep registration of the patient's previous scan trims registration from 90 s to 4.1 s and lifts SSIM, PSNR, and NMSE.","key_machinery":"The mechanism is a three-stage pipeline: an initial reconstruction network (E2E-VarNet) maps undersampled multi-coil k-space data to an image; a deep registration model (EasyReg) non-linearly aligns the prior subject-specific scan to that initial reconstruction; and a transformer-based enhancement network, using multi-head self-attention with the registered prior injected through shift and scale operations, produces the final image. The registered prior is patched and embedded like a Vision Transformer input, and its embeddings modulate the self-attention outputs. Training uses an SSIM loss between the enhanced reconstruction and the fully sampled reference. The paper's key argument is that each replacement -- deep registration for linear registration, transformer for UNet -- drives both the quality gain and the 90-second to 4.1-second speedup.","core_discovery":"The central claim is that replacing linear registration with a deep-learning registration model and replacing a UNet enhancer with a transformer-based enhancer improves the quality of prior-informed MRI reconstruction at every tested acceleration factor while making the method much faster. The pipeline first reconstructs the undersampled scan with E2E-VarNet, then registers the patient's previous scan to that reconstruction using a fast learning-based registration model, and finally feeds both the initial reconstruction and the registered prior into a transformer enhancement network whose self-attention and shift/scale conditioning refine the image. On a held-out test set of 18 subjects, the transformer-enhanced version reports the highest SSIM and PSNR and the lowest NMSE at R5, R10, R15, and R20, and the advantage over the non-enhanced baseline widens as acceleration increases. The same enhanced reconstructions also yield higher Dice coefficients and lower absolute volume errors when segmented with an automated tool, supporting the claim that the improvement is not merely cosmetic.","pith_inferences":["A natural extension is to test the pipeline in patients with new pathology between scans, such as new multiple sclerosis lesions; the current healthy-subject dataset cannot show whether small clinically relevant lesions survive the enhancement.","If the reported gains transfer to other contrasts and field strengths, prior-informed reconstruction could become a general acceleration strategy, with the deep registration speed making it feasible to run at the scanner.","The atlas ablation suggests a fallback when no previous scan exists: a registered population template may provide part of the benefit, though less than a true subject-specific prior; a direct comparison on the same subjects would quantify that gap.","Because the improvement grows with acceleration factor, the method's value may be largest for very aggressive subsampling beyond R20, where motion and scan time are most reduced; evaluating R25 or R30 would show whether the trend continues."],"forward_implications":["At every acceleration factor from R5 to R20, the transformer-enhanced method reports higher SSIM and PSNR and lower NMSE than both non-enhanced reconstructions and the linearly registered UNet-enhanced baseline, with the gap widening at higher acceleration.","Replacing linear registration with the deep registration model cuts registration time from roughly 90 seconds to 4.1 seconds per subject, removing a major bottleneck to using prior scans in a clinical workflow.","The enhanced reconstructions lead to better downstream brain segmentation, with higher Dice similarity coefficients and smaller absolute volume errors for the cortex, white matter, and hippocampus compared with non-enhanced reconstructions.","Using a population atlas as the prior instead of the subject's own previous scan yields worse reconstruction quality, indicating that subject-specific prior information is the source of the gain.","Blinded expert readers rated the enhanced reconstructions higher or equal to non-accelerated images on overall quality, contrast, and fine anatomical detail, suggesting that the acceleration does not visibly degrade diagnostic quality."],"supporting_citations":[{"why":"Supplies the prior-informed reconstruction framework and the linearly registered UNet-enhanced baseline that this paper extends and outperforms.","marker":"[12]"},{"why":"Supplies the E2E-VarNet model used for the initial reconstruction of undersampled k-space data.","marker":"[17]"},{"why":"Supplies the deep-learning registration model that replaces linear registration and gives the reported speed-up.","marker":"[18]"},{"why":"Supplies the linear registration method whose approximate 90-second runtime is the baseline timing comparison.","marker":"[20]"},{"why":"Supplies the Vision Transformer patch-and-embed approach used by the transformer-based enhancement network.","marker":"[14]"},{"why":"Supplies the multi-head self-attention mechanism on which the enhancement network is built.","marker":"[15]"},{"why":"Supplies the SynthSeg tool used to measure downstream brain segmentation quality on the reconstructed images.","marker":"[21]"}],"fun_headline_variants":["Deep registration speeds MRI recon 22x while improving quality","Transformer-enhanced prior-informed MRI wins on quality and speed","Prior scans boost MRI reconstruction with deep-reg and transformers","Faster, sharper MRI: deep registration plus transformer enhancer","MRI reconstruction improved at R5-R20 using prior scans and AI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the linearly registered UNet-enhanced baseline was trained and evaluated under the same conditions as the proposed method; if the baseline numbers came from a differently split or preprocessed run, the reported improvement could be an artifact of experimental setup rather than the new architecture.","fun_headline_variants_meta":{"raw":{"variants":["Deep registration speeds MRI recon 22x while improving quality","Transformer-enhanced prior-informed MRI wins on quality and speed","Prior scans boost MRI reconstruction with deep-reg and transformers","Faster, sharper MRI: deep registration plus transformer enhancer","MRI reconstruction improved at R5-R20 using prior scans and AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00022,"raw_usage":{"total_tokens":1475,"prompt_tokens":1000,"completion_tokens":475,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":616,"completion_tokens_details":{"reasoning_tokens":393}},"tokens_in":616,"tokens_out":475,"duration_ms":6492,"temperature":1.0,"reasoning_tokens":393,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:51:11.503424+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the previous linearly registered UNet-enhanced method on the exact same training, validation, and test split, preprocessing, and hardware used here, then compare SSIM, PSNR, and NMSE on the same test subjects; if the proposed method no longer wins the Wilcoxon signed-rank test at p < 0.05, the central claim fails. A second check is whether expert readers can still detect the improved reconstructions as anatomically faithful in patients with new lesions.","supporting_citations":[{"cited_title":"Souza, Y","cited_arxiv_id":null,"evidence_quote":"Supplies the prior-informed reconstruction framework and the linearly registered UNet-enhanced baseline that this paper extends and outperforms."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the deep-learning registration model that replaces linear registration and gives the reported speed-up."},{"cited_title":"Vaswani, N","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-head self-attention mechanism on which the enhancement network is built."}],"review_version":1}