{"id":"8f22832c-2ed8-4046-91fb-8f8ab14c8ae4","arxiv_id":"1908.09262","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Recon-GLGAN adds a global-local context discriminator to a GAN, using a heart ROI as extra training signal, and reports small reconstruction improvements on ACDC cardiac MRI.","lead":"The authors propose a GAN for cardiac MRI reconstruction whose discriminator looks at both the full image and a heart region defined by segmentation masks. The method gives small metric gains over baseline GANs, but the reported 'significant' improvement is not backed by significance tests.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline claim of improvement across all metrics is contradicted by the paper's own tables, and no significance testing supports the word 'significant'.","rationale":"The reader's verdict of CONDITIONAL is appropriate, and I agree that the paper needs revision before the central claim can be accepted. However, the single most load-bearing concern is not the train/test ROI mismatch identified by the reader. The ROI mismatch is a reasonable generalizability worry, but it is speculative and mostly affects the ROI-specific numbers and downstream segmentation. The decisive issue is that the paper's own reported numbers contradict the strongest wording of the central claim. Table 1 already contains entries where Recon-GLGAN is numerically worse than GAN, and Table 3 contains entries where the GL variant is worse than the corresponding baseline. These are internal, checkable inconsistencies rather than assumptions about external data. Moreover, the effect sizes are far smaller than the reported standard deviations, and no significance test is provided, so there is no statistical evidence for the word 'significant'. The method itself is plausible, the code is available, and the idea of injecting an ROI prior through the discriminator is reasonable, so this should not be a rejection. A conditional acceptance that requires the authors to add paired significance testing, correct any transcription errors, and soften the claim from 'all metrics and significant' to 'most metrics on average' is the right outcome. The reader's verdict already contains this condition, so I leave it unchanged.","tokens_in":8933,"tokens_out":7273,"duration_ms":74107,"concrete_test":"Re-run the evaluation on the 1076 test slices in paired fashion. For every slice and every cell in Tables 1 and 3, compute the per-slice difference (Recon-GLGAN or GL variant minus the corresponding baseline) for each metric, acceleration factor, and ROI, then report the mean difference, a 95% bootstrap confidence interval, and a paired Wilcoxon signed-rank p-value. Also count how many of the 1076 slices favor the proposed method per metric. If any claimed improvement has a confidence interval crossing zero, a p-value above 0.05, or a majority of per-slice differences going the wrong way, the 'across all metrics' and 'significant' claims must be weakened. The same check will reveal whether the point-estimate reversals in Table 1 (2x ROI NMSE, 8x ROI SSIM) and Table 3 (GL-ReconGAN and GL-DAGAN) are actual results or transcription errors.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, stated in Section 3.4 as performing better than baseline GAN and U-Net across all metrics for all acceleration factors, is not supported by the numbers in Tables 1 and 3. In Table 1, two entries go the wrong way: at 2x acceleration in the ROI, Recon-GLGAN has NMSE 0.00944 versus GAN's 0.00942, and at 8x in the ROI, Recon-GLGAN has SSIM 0.7293 versus GAN's 0.7294. Table 3, which is meant to show that the context discriminator can be dropped into existing GAN methods, also contains reversals: GL-ReconGAN's ROI NMSE and PSNR are worse than ReconGAN's (0.01836 vs 0.01800 and 25.72 vs 25.76), and GL-DAGAN is worse than DAGAN on full-image SSIM (0.8450 vs 0.8452) and on ROI NMSE and PSNR (0.01858 vs 0.01838 and 25.62 vs 25.68). The reported standard deviations are on the order of 0.01 in NMSE and 2.7-3.6 dB in PSNR, while the claimed advantages are often 0.01-0.1 dB, so without paired significance tests the abstract's 'significant performance improvement' is unjustified. This is an internal inconsistency between the paper's strongest sentence and the measurements it reports, not an external assumption about annotation protocols.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Recon-GLGAN, a GAN-based MRI reconstruction method with a U-Net generator and a context discriminator that combines a global feature extractor on the full image and a local feature extractor on a 60x60 region of interest (ROI) derived from cardiac segmentation masks. The authors evaluate on the ACDC cardiac MRI dataset at 2x, 4x, and 8x acceleration, reporting NMSE, PSNR, and SSIM for both full image and ROI. They further claim that the context discriminator can be dropped into existing GAN-based reconstruction methods (ReconGAN, SEGAN, ComGAN, DAGAN) to improve them, and they present a preliminary segmentation experiment suggesting that reconstructions from their method yield segmentation results similar to fully sampled images. The paper's central claims are that the proposed method outperforms baseline GAN and U-Net across all metrics and acceleration factors, and that the ROI prior improves reconstruction and downstream segmentation.","tokens_in":9192,"tokens_out":6518,"duration_ms":60069,"significance":"If the claims were fully supported, the paper would make a useful contribution by demonstrating that a task-driven ROI prior, extracted from segmentation masks and injected through the discriminator, can improve GAN-based MRI reconstruction and segmentation. The idea is relevant to application-driven MRI and the code release is a positive factor for reproducibility. The controlled comparison in Table 3, where only the generator loss is varied while the generator and discriminator design are fixed, is a reasonable experimental design for isolating the effect of the loss terms. However, the current evidence does not establish the headline claims: the paper's own tables contain counterexamples to the 'across all metrics' statement, no statistical significance tests are provided, the extension to existing methods is not tested with their original architectures, and the segmentation evidence is based on only three samples with a proxy ground truth. The central concept may be sound, but the manuscript requires substantial revision before the claims can be accepted.","major_comments":[{"comment":"The sentence 'Our model Recon-GLGAN performs better than the baseline GAN and U-Net across all metrics for all acceleration factors' is contradicted by the paper's own measurements: at 2x in the ROI, Recon-GLGAN NMSE (0.00944) is worse than GAN (0.00942), and at 8x in the ROI, Recon-GLGAN SSIM (0.7293) is worse than GAN (0.7294). The claim should be revised to report the direction and magnitude of improvements per metric and acceleration factor, acknowledging these exceptions.","section":"Section 3.4, Table 1"},{"comment":"The term 'significant' is used repeatedly ('significant performance improvement', 'significantly better reconstruction performance') but no statistical significance tests are reported. The reported standard deviations (e.g., NMSE ±0.01, PSNR ±2.7–3.6 dB) are of the same order as or larger than the mean differences (often 0.01–0.1 dB in PSNR), so the observed differences could well arise from noise. Paired significance tests across test slices (e.g., Wilcoxon signed-rank test) are required to support the word 'significant'.","section":"Abstract, Section 3.4, Tables 1 and 3"},{"comment":"The claim that the context discriminator 'can be extended to existing GAN based reconstruction architectures' is not supported by the experimental design: for all methods (ReconGAN, SEGAN, ComGAN, DAGAN) the generator is replaced by U-Net and the discriminator is either a basic global classifier or the context discriminator, so only the loss terms remain from the original works. This tests adding the context discriminator to a U-Net-based GAN with different losses, not the original architectures. Moreover, Table 3 contains reversals (GL-ReconGAN ROI NMSE 0.01836 vs ReconGAN 0.01800 and ROI PSNR 25.72 vs 25.76; GL-DAGAN full-image SSIM 0.8450 vs DAGAN 0.8452, ROI NMSE 0.01858 vs 0.01838), so the statement that context discriminator variants 'have shown improved results' is not universally true.","section":"Section 3.4, Table 3"},{"comment":"The training-time ROI is defined as the 60x60 bounding box around the segmentation mask, whereas the test-time ROI is obtained from the algorithm in [7]. No evidence is provided that these two ROI definitions identify the same heart region across test slices. Because the local feature extractor is trained only on mask-derived ROIs, the reported ROI improvements evaluated on [7] ROIs may not generalize if the definitions diverge. Report the overlap (e.g., Dice or IoU) between the two ROIs on the test set, or evaluate ROI metrics under both definitions.","section":"Sections 3.1 and 3.3"},{"comment":"The segmentation demonstration is based on only three test images, uses the U-Net output on the fully sampled image as a proxy for ground truth because test masks are unavailable, and provides no error bars or statistical comparison across the 1076 test slices. This evidence is too limited to support the abstract's statement that 'the reconstructions from the proposed method give segmentation results similar to fully sampled images.' Either restrict the claim to a qualitative illustration for the three samples or evaluate segmentation with a dataset that provides test masks.","section":"Section 3.4, Figure 6"}],"minor_comments":[{"comment":"The first sentence contains a punctuation error: 'In this work. we proposed' should read 'In this work, we proposed.'","section":"Section 4"},{"comment":"The phrase 'The dimensions of the ROI is set' has a subject-verb agreement error; it should be 'are set.'","section":"Section 3.1"},{"comment":"The naming is inconsistent: the architecture is called 'ComGAN' in the text and 'COMGAN' in Table 2; please standardize the name.","section":"Tables 2 and 3"},{"comment":"The notation pG(xu) in the second expectation is undefined and likely a typo; the expectation should be over the distribution of under-sampled inputs, e.g., x_u ~ p_data(x_u).","section":"Section 2.3, Eq. (1)"},{"comment":"The description of the baseline GAN in Table 1 is not explicit; please state clearly that the baseline GAN uses the same generator and a global-only discriminator (i.e., the context discriminator without the local path), or describe the exact baseline architecture in the text.","section":"Section 2.4"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea of using a segmentation-mask-derived ROI prior in the discriminator is worth exploring, and the code release is a positive factor. However, the manuscript currently overstates its results: the 'across all metrics' claim is internally contradicted by the tables, 'significant' is used without statistical tests, and the extension and segmentation claims are only weakly evidenced. A careful revision with corrected claims, added significance testing, and a more robust segmentation evaluation would be needed before acceptance. The novelty is moderate given the earlier global-local discriminator in image inpainting (Iizuka et al.), but the application to MRI reconstruction with a task-driven ROI prior could still be a useful contribution if the evidence is cleaned up."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nI agree with the conditional verdict. This is a plausible ROI-aware GAN for cardiac MRI reconstruction, and the code is available. The core transfer — taking Iizuka et al.'s global-local discriminator and attaching a 60×60 ROI derived from training segmentation masks — is legitimate, and the drop-in experiment on other GAN losses is a useful test of the concept. The architecture is clean and the paper is easy to follow.\n\nThe problem is the headline. The abstract and Section 3.4 claim significant improvement across all metrics, but the numbers in Table 1 do not support that. At 2x ROI NMSE, Recon-GLGAN is 0.00944 vs GAN's 0.00942; at 8x ROI SSIM, 0.7293 vs 0.7294. Those are reversals, not ties. Table 3 has similar reversals: GL-ReconGAN's ROI NMSE and PSNR are worse than ReconGAN's, and GL-DAGAN is worse than DAGAN on full-image SSIM and on ROI NMSE/PSNR. Most reported gains are within the standard deviations, and no significance tests are presented. 'Significant' is doing work the statistics cannot support.\n\nThe segmentation experiment is preliminary: three samples, no error bars. It is fine as a demonstration, not as a claim that reconstructions give FS-equivalent segmentation.\n\nOne extra soft spot: the training ROI is centered on the segmentation mask bounding box, while the test ROI comes from a different algorithm. If those localizations drift, the local discriminator may be trained on a region that the evaluation does not reward. The paper should check this directly.\n\nNone of this kills the paper. The aggregate direction favors Recon-GLGAN, and the extensibility result is worth having. But the central claim needs reworded, paired significance tests need to be run, and the ROI mismatch needs a paragraph. That is a solid conditional-accept path.\n\nYes, this deserves a serious referee. I would cite it as a reference for ROI-aware GAN reconstruction, though I would not trust the headline numbers without rechecking.","headline":"A plausible ROI-aware GAN for MRI reconstruction whose aggregate direction is positive but whose headline claim outruns its own tables.","tokens_in":9782,"tokens_out":2485,"would_cite":true,"duration_ms":25118,"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":"A GAN whose discriminator looks at both the full image and a heart-region crop reconstructs undersampled cardiac MRI more accurately than baseline GAN and U-Net, and the same discriminator can be dropped into other reconstruction GANs to…","keywords":["cardiac MRI","MRI reconstruction","generative adversarial network","global-local context","region of interest","segmentation","deep learning","undersampled k-space"],"falsifier":"Run the same training protocol but with a fixed central 60x60 crop as the local input at both train and test; if the segmentation-prior version does not beat this fixed-crop version on held-out slices, the claim that mask-derived ROIs drive the gain is falsified.","tokens_in":8705,"feed_emoji":"🧲","tokens_out":6936,"duration_ms":64671,"temperature":0.7,"pith_summary":"The paper sets out to show that a GAN reconstructing undersampled cardiac MRI improves when its discriminator is given two views: the whole image and a crop around the heart. This Recon-GLGAN model uses a U-Net generator and a context discriminator with global and local feature paths, and it is trained with L1 plus adversarial loss. The authors report better NMSE, PSNR, and SSIM than baseline GAN and U-Net at 2x, 4x, and 8x acceleration, better performance when the same discriminator replaces the standard one in other GAN reconstructions, and segmentation outputs close to those from fully sampled images. A sympathetic reader would take the paper's claim to be that application-driven region-of-interest priors are a useful, portable ingredient for MRI reconstruction GANs.","feed_headline":"GAN with a global-plus-local discriminator beats MRI baselines","feed_subtitle":"Adding a heart-region crop to the discriminator improves reconstruction at 2x, 4x, and 8x acceleration.","key_machinery":"The central object is the context discriminator: a discriminator with two parallel feature extractors whose outputs are concatenated and classified. The global feature extractor sees the entire image; the local feature extractor sees a 60x60 region of interest that, during training, is centered on the bounding box of the heart's segmentation mask. The concatenated 128-dimensional feature vector is passed through one fully connected layer with a sigmoid to yield the real/fake probability. This two-path design carries the argument because it makes the adversarial loss reward reconstructions that are locally convincing inside the clinical region of interest as well as globally coherent, and it is the component the paper swaps into other GAN reconstruction pipelines to show portability.","core_discovery":"On its own terms, the paper claims that reconstruction quality is governed less by generator capacity than by where the discriminator directs attention. The context discriminator concatenates a 64-dimensional feature vector from a global path over the full 160x160 image with a 64-dimensional vector from a local path over a 60x60 heart-region crop, and a single fully connected layer classifies the resulting 128-dimensional vector as real or fake. Because the local crop at training time is centered on each slice's segmentation-mask bounding box, the generator is pressured to reconstruct the diagnostically relevant region faithfully while the global path maintains whole-image consistency. On the ACDC cardiac MRI data, the paper reports that this design beats the baseline GAN and U-Net on full-image and ROI metrics for 2x, 4x, and 8x accelerations, that retrofitting the discriminator into ReconGAN, SEGAN, ComGAN, and DAGAN improves most of their 4x metrics, and that its reconstructions yield segmentation Dice and Hausdorff scores close to those of fully sampled images.","pith_inferences":["A natural extension the paper does not test is whether the same ROI-prior design helps other organs or imaging contrasts when the region of interest is defined by an automatic localizer; if it does, the principle is general, not cardiac-specific.","The training-time ROI from segmentation masks and the test-time ROI from a separate localization algorithm are not guaranteed to coincide, so a robustness study that perturbs the test crop or trains with the same localizer at both stages would reveal how dependent the gains are on this assumption.","The reported improvements are small in absolute metric terms; deciding whether they are clinically meaningful would require task-based tests such as ejection-fraction estimation or expert reading rather than pixel-similarity numbers alone."],"forward_implications":["The context discriminator can be treated as a plug-in replacement for the discriminator in other GAN-based MRI reconstruction methods, so an existing pipeline can gain ROI-aware behavior without redesigning its generator or loss.","Gains are larger for 4x and 8x acceleration than for 2x, so the ROI prior matters most exactly when aliasing is hardest to remove.","Because the local path reduces reconstruction error in the heart region more than in the whole image, the network is effectively trading global capacity for diagnostic-region fidelity.","Using Recon-GLGAN reconstructions instead of zero-filled or baseline-GAN images as input to a pre-trained segmentation U-Net gives Dice and Hausdorff scores closer to fully sampled input, so the reconstruction benefit propagates to a downstream segmentation task."],"supporting_citations":[{"why":"Supplies the ACDC cardiac MRI training and test slices with segmentation masks used for ROI extraction and evaluation.","marker":"[1]"},{"why":"Provides the global-plus-local discriminator design that the context discriminator is adapted from.","marker":"[5]"},{"why":"Supplies the conditional adversarial training formulation underlying the generator-discriminator setup.","marker":"[6]"},{"why":"Provides the algorithm used to obtain test-set regions of interest for evaluating reconstruction quality.","marker":"[7]"},{"why":"Serves as one of the GAN reconstruction baselines whose discriminator is replaced to test portability.","marker":"[10]"},{"why":"Defines the U-Net architecture used as the generator and as the segmentation network in downstream evaluation.","marker":"[12]"},{"why":"Provides the application-driven MRI motivation and the idea of exploiting undersampled-data segmentation priors.","marker":"[13]"},{"why":"Serves as another GAN reconstruction baseline used in the discriminator-replacement experiment.","marker":"[15]"}],"fun_headline_variants":["Global-local discriminator sharpens MRI at 2x-8x acceleration","Heart-region focus improves GAN-based MRI reconstruction","Context discriminator lifts GAN MRI quality and segmentation","Regional attention in GAN boosts MRI reconstruction speed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training-time heart region is defined by segmentation-mask bounding boxes while test-time regions come from a different algorithm, and the whole region-of-interest benefit rests on these two identifying the same region on every slice.","fun_headline_variants_meta":{"raw":{"variants":["Global-local discriminator sharpens MRI at 2x-8x acceleration","Heart-region focus improves GAN-based MRI reconstruction","Context discriminator lifts GAN MRI quality and segmentation","Regional attention in GAN boosts MRI reconstruction speed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000209,"raw_usage":{"total_tokens":1446,"prompt_tokens":1021,"completion_tokens":425,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":637,"completion_tokens_details":{"reasoning_tokens":358}},"tokens_in":637,"tokens_out":425,"duration_ms":4572,"temperature":1.0,"reasoning_tokens":358,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:16:09.610439+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same training protocol but with a fixed central 60x60 crop as the local input at both train and test; if the segmentation-prior version does not beat this fixed-crop version on held-out slices, the claim that mask-derived ROIs drive the gain is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the ACDC cardiac MRI training and test slices with segmentation masks used for ROI extraction and evaluation."},{"cited_title":"ACM Trans","cited_arxiv_id":null,"evidence_quote":"Provides the global-plus-local discriminator design that the context discriminator is adapted from."},{"cited_title":"In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)","cited_arxiv_id":null,"evidence_quote":"Supplies the conditional adversarial training formulation underlying the generator-discriminator setup."},{"cited_title":"Medical Image Analysis 51, 21 – 45 (2019)","cited_arxiv_id":null,"evidence_quote":"Provides the algorithm used to obtain test-set regions of interest for evaluating reconstruction quality."},{"cited_title":"IEEE Transactions on Medical Imaging 38(1), 167–179 (Jan 2019)","cited_arxiv_id":null,"evidence_quote":"Serves as one of the GAN reconstruction baselines whose discriminator is replaced to test portability."},{"cited_title":"In: Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015","cited_arxiv_id":null,"evidence_quote":"Defines the U-Net architecture used as the generator and as the segmentation network in downstream evaluation."},{"cited_title":"In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2018","cited_arxiv_id":null,"evidence_quote":"Provides the application-driven MRI motivation and the idea of exploiting undersampled-data segmentation priors."},{"cited_title":"IEEE Transactions on Medical Imaging 37(6), 1310–1321 (June 2018)","cited_arxiv_id":null,"evidence_quote":"Serves as another GAN reconstruction baseline used in the discriminator-replacement experiment."}],"review_version":1}