{"id":"061e568c-595b-49d0-862f-267c50d529c7","arxiv_id":"2507.00670","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SDR generates semantically diverse, data-consistent MRI reconstructions that improve automated pathology detection recall compared to standard or resampled reconstructions, without hurting precision.","lead":"The authors show that accelerated MRI reconstruction methods, including diffusion-based ones, can miss small, rare pathologies. They introduce SDR, a post-hoc method that produces multiple data-consistent reconstructions with greater semantic diversity, and report that an automated detector finds more true pathologies without more false positives.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central clinical claim rests on a detector proxy whose training labels are shared with the SDR objective; the reported recall gains may reflect optimization against the evaluation metric rather than recovery of true missed pathology.","rationale":"The reader's weakest assumption identifies the detector-as-proxy issue, and I agree that this is the most load-bearing concern. The paper is otherwise clearly written, the method is specified in detail, and the code is released, so reproducibility is not the main issue. The central risk is that both the generation objective and the evaluation metric are derived from the same fastMRI+ label set: SDR optimizes feature distances in a ViTDet trained on those labels, and the 'independent' Faster-RCNN is trained on the same labels. This makes the measured recall improvement vulnerable to shortcut learning or reward hacking, where the reconstructed images become more detectable to a network trained on the same distribution without becoming more clinically faithful. Data consistency does not fully prevent this, since many images satisfy the measured k-space data. The missing random-perturbation control is a related but secondary concern: even if SDR outperforms it, the clinical relevance of the detected boxes still requires human validation. Therefore the verdict remains conditional, with the condition being a clinical reader study or equivalent independent validation.","tokens_in":7415,"tokens_out":5826,"duration_ms":82452,"concrete_test":"Conduct a blinded radiologist reader study on a random subset of fastMRI+ test knees (e.g., 100 volumes): each reader independently marks the 10 pathology classes on (a) the original reconstruction and (b) the SDR set, presented in random order without method labels, and score against the fastMRI+ ground-truth boxes using recall and false-positive rate. If SDR does not improve radiologist-level recall, or if it increases false positives, the clinical claim is unsupported; if radiologists do find more true pathologies in the SDR set, the shared-label circularity concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's key quantitative claim—that SDR 'capture[s] more possible pathologies' and reduces false-negative diagnoses—is measured entirely by a separate Faster-RCNN detector trained on fastMRI+ annotations (Section 4). The SDR objective (Eq. 3) maximizes feature distance in a ViTDet backbone that was itself fine-tuned on the same fastMRI+ annotations (Sections 3.3-3.4). Thus the evaluation detector is not causally independent of the generative objective: both share the same label distribution. Because accelerated MRI is ill-posed, the data-consistency projection DC and the ℓ2-ball constraint (r=3) still permit large null-space changes; optimizing a detector-trained feature space can reshape image content to trigger pathology detectors without necessarily recovering anatomy a radiologist would call present. The mAP0.25 and recall improvements in Fig. 4 could therefore be a form of reward hacking on the shared annotation distribution. No reader study, independent annotation source, or clinician validation is provided, so the abstract's wording about false-negative diagnoses exceeds what the current evidence supports. A random-perturbation control is also missing, but the deeper issue is that the evaluation signal and the generative objective are not sufficiently separated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses the clinical risk that accelerated MRI reconstruction may miss small or rare pathologies. It proposes Semantically Diverse Reconstructions (SDR), a post-hoc method that takes an initial reconstruction and generates additional reconstructions that are data-consistent but semantically different in regions proposed by an object detector or by a user. Semantic diversity is measured as feature distance in a ViTDet backbone that is adversarially fine-tuned for robust gradients. The method is evaluated on fastMRI+ knee data with three base reconstruction techniques (DDIP, CSGM, E2E-VarNet), using a separately trained Faster-RCNN detector as a proxy for clinical reading. The reported results show higher recall and comparable or better mAP0.25 for SDR compared with single reconstructions and with repeated sampling from diffusion models.","tokens_in":7675,"tokens_out":2910,"duration_ms":36586,"significance":"The problem addressed is important: reconstruction quality metrics do not capture whether clinically relevant information is preserved, and false-negative diagnoses are a concrete risk in accelerated MRI. The proposed method is interesting because it is post-hoc, applicable to any reconstruction method, and relatively fast, and the authors provide code. The paper also makes a useful conceptual contribution by shifting evaluation from image quality toward pathology detection. However, the current evidence is not yet sufficient to support the strong clinical claims in the abstract. The main quantitative comparison rests on a single bar chart without error bars or significance tests, the generative objective and the evaluation detector are trained on the same fastMRI+ label distribution, and there is no ablation against random data-consistent perturbations. If these gaps are addressed, the contribution could be valuable to the medical imaging community.","major_comments":[{"comment":"The central quantitative claim that SDR 'significantly reduces' false negatives is not supported by the reported statistics. Fig. 4 shows single bars without error bars, confidence intervals, or significance tests. Please report mean and variance across subjects or repeated runs and provide a paired statistical test (e.g., Wilcoxon signed-rank) for recall and mAP0.25 between SDR and each baseline.","section":"Section 4, Fig. 4"},{"comment":"There is a circularity concern between the generative objective and the evaluation metric. The semantic distance in Eq. (3) is computed in the feature space of a ViTDet fine-tuned on fastMRI+, and the success metric is the recall of a Faster-RCNN trained on the same fastMRI+ annotations. Even though the two detectors are architecturally different, they share the same label distribution and are trained on the same pathology annotations. Thus the reported recall gains may reflect optimization against the evaluation signal rather than recovery of pathology that a clinician would confirm. Please provide an evaluation with an independent annotation source (e.g., a small reader study or a held-out annotation set) or otherwise demonstrate that the gains persist when the evaluation signal is not the training signal of the feature encoder.","section":"Sections 3.3, 3.4, and 4"},{"comment":"No ablation against random data-consistent perturbations is reported. Because the method projects onto an ℓ2 ball of radius r=3 around the initial reconstruction and then applies data consistency, a necessary control is to generate the same number of reconstructions with random perturbations inside the same ball, followed by the same data-consistency projection, and to measure recall and mAP. Without this control, the specific contribution of the semantic diversity objective in Eq. (2) cannot be isolated from the effect of merely exploring a set of data-consistent solutions.","section":"Section 3.2 and Fig. 4"},{"comment":"The evaluation equates object-detector recall with a reduction in false-negative diagnoses. The paper itself states that clinical assessment would ideally require clinician input, but no clinical validation is provided. The abstract's phrasing about 'false-negative diagnoses' exceeds what the current evidence supports. Please either temper the clinical claims or add evidence that the detector's detections correspond to radiologist-identified pathologies, for example by reporting per-class recall and a qualitative evaluation on a small clinician-annotated subset.","section":"Section 4"}],"minor_comments":[{"comment":"The variable Nopt appears in Algorithm 1 but is not defined in the text; please define it and clarify the relationship between Eq. (2) and the initialization of each reconstruction with Gaussian noise.","section":"Algorithm 1 and Section 3.2"},{"comment":"The name of the baseline is written inconsistently as 'E2E-VarNet' and 'E2E-Varnet'; please unify the spelling.","section":"Throughout"},{"comment":"The bar chart does not show numerical values, which makes it difficult for readers to assess the magnitude of the reported differences; consider adding value labels or a corresponding table.","section":"Fig. 4"},{"comment":"The caption states that SDR 'generates clinically relevant reconstructions outperforming resampling', which is stronger than the evidence presented; please soften the wording to match the statistical support.","section":"Fig. 4 caption"},{"comment":"Reference [5] is cited as evidence that diffusion models rarely sample less likely but diagnostically relevant solutions; a sentence elaborating on this finding would help the reader understand the motivation.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of the venue and addresses a timely problem. The main reason for major revision rather than rejection is that the central idea is plausible and the identified issues are addressable within the manuscript's scope: adding statistical rigor, a random-perturbation control, and an independent evaluation signal would substantially strengthen the claims. I would also encourage the editor to verify that the authors make the trained evaluation detector and the exact data splits available, since the circularity concern can only be assessed by the community if the evaluation pipeline is fully reproducible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi [Colleague],\n\nYou should know about this one: it proposes a genuinely new post-hoc method, SDR, that takes any accelerated MRI reconstruction and generates several additional reconstructions that are all data-consistent but semantically different, specifically in regions where pathologies might lurk. The trick is to maximize feature distance in a robust ViTDet backbone, and the authors show this can uncover pathologies missed by the original reconstruction or by repeated sampling from diffusion models. The idea is clever, the method is clearly described, they ship code, and it runs about 3 seconds per image, which makes it practical.\n\nWhat’s actually new: previous work either didn’t generate multiple reconstructions or, like diffusion models, sampled from the posterior but rarely hit the rare pathologies. SDR deliberately pushes reconstructions in semantically distinct directions, and the use of an adversarially robust detector for semantic distance is a sensible move. The manual and automatic proposal modes are a nice touch.\n\nThe soft spots are all in the evaluation. The main evidence is detector recall and mAP from a Faster-RCNN trained on fastMRI+ pathology labels, but the SDR objective itself uses a ViTDet fine-tuned on those same labels. So the generator and evaluator share a training distribution, and the measured recall gains could partly reflect optimizing toward that label distribution rather than toward anatomy a radiologist would flag. The paper calls the evaluation \"independent\" because a different architecture is used, but that doesn’t establish causal independence. This isn’t a fatal flaw—it’s a serious confound that needs addressing with either a reader study, an independent annotation source, or at least a demonstration that random data-consistent perturbations don’t achieve similar recall. That control is missing, which is a shame because it would isolate the contribution of the semantic objective.\n\nAlso, no error bars, confidence intervals, or significance tests appear anywhere, yet the abstract says \"significantly reduces.\" That’s unsupported. The qualitative examples are suggestive but not convincing on their own.\n\nThe comparison baselines are okay but not exhaustive; they don’t compare to other diversity-focused methods, like the one from Cohen et al. [5] they cite.\n\nSo: the core algorithmic idea is interesting and likely useful for uncertainty visualization, but the clinical claim—reducing false-negative diagnoses—exceeds the evidence. The paper deserves a careful peer review, but it needs major experimental revision before I’d trust the headline result. If you’re working on MRI reconstruction or uncertainty quantification, it’s worth reading for the method; don’t cite it yet as evidence of clinical benefit.","headline":"Smart idea for semantically diverse MRI reconstructions, but the detector-based evaluation shares labels with the objective and lacks statistical support, so the clinical claim is overreaching.","tokens_in":8172,"tokens_out":3701,"would_cite":true,"duration_ms":42606,"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":"Producing multiple semantically different, data-consistent MRI reconstructions reveals pathologies that a single accelerated reconstruction misses.","keywords":["accelerated MRI","MRI reconstruction","pathology detection","false negatives","semantic diversity","data consistency","adversarial robustness","diffusion models"],"falsifier":"Show the fastMRI+ test slices where SDR recovers a pathology that the baseline reconstruction missed to a panel of radiologists, without telling them which reconstruction came from SDR, and ask them to mark the pathologies they can see; if the radiologists do not see the additional pathologies, the claim that SDR uncovers clinically relevant image details is falsified.","tokens_in":7219,"feed_emoji":"🧲","tokens_out":8509,"duration_ms":85817,"temperature":0.7,"pith_summary":"The paper's goal is to show that accelerated MRI reconstruction, even with modern diffusion-based methods, can silently drop small or rare pathologies, and that this risk is reduced by examining a set of reconstructions instead of a single one. It introduces Semantically Diverse Reconstructions (SDR), a post-hoc procedure that takes any initial reconstruction and produces additional reconstructions which are forced to be semantically different in candidate pathology regions while remaining fully consistent with the measured k-space data. On the fastMRI+ knee dataset, using a separately trained Faster-RCNN detector, SDR reconstruction sets detect more true pathologies (higher recall) than the original reconstructions, repeated diffusion sampling, or E2E-VarNet ensembles, and mean average precision does not drop but improves. If the result holds, SDR offers a practical route to surfacing diagnostic uncertainty in accelerated MRI without changing the acquisition.","feed_headline":"Diverse MRI reconstructions surface pathologies single scans miss","feed_subtitle":"SDR's extra reconstructions find pathologies that standard MRI reconstructions miss, in about three seconds per image.","key_machinery":"The central mechanism is an iterative projected-gradient loop that maximizes the semantic distance $d(\\hat{x}^{(i)}, \\hat{x}^{(j)}) = \\sum_{b \\in \\mathcal{B}} \\| f_\\phi(\\hat{x}^{(i)}, b) - f_\\phi(\\hat{x}^{(j)}, b) \\|_2$ over a set of proposal boxes $\\mathcal{B}$, where $f_\\phi$ is a box-feature encoder built on a ViTDet detector with an adversarially robust ViT backbone. Each update is followed by a projection $P_{\\mathcal{B}(\\hat{x}^{(1)}, r)}$ onto the $\\ell^2$-ball of radius $r$ around the initial reconstruction and a data-consistency projection that substitutes measured k-space values; the robust backbone is what turns the feature-space gradient into a semantically meaningful image change. Proposal boxes come either from the ViTDet region proposal network (SDR-A) or from a manual annotation (SDR-M).","core_discovery":"The paper demonstrates a method rather than a new theory: given an initial reconstruction $\\hat{x}^{(1)}$ and a set of candidate bounding boxes in likely pathology regions, SDR generates $N_\\mathrm{rec}-1$ additional reconstructions by optimizing each new image so that the $\\ell^2$ distance between the ViTDet box features of that image and the features of all previously generated reconstructions is maximized. A projection onto the $\\ell^2$-ball of radius $r$ around $\\hat{x}^{(1)}$ and a data-consistency step that re-inserts measured k-space values keep every reconstruction both close to the original and fully consistent with the measurement. The paper reports that an adversarially robust ViT backbone is required for these feature gradients to be semantically meaningful, and that the resulting reconstruction set achieves higher pathology recall than repeated sampling of the diffusion-based DDIP and CSGM baselines or retrained E2E-VarNet ensembles, while also improving mean average precision.","pith_inferences":["The same recipe -- maximize distance in a robust feature space under data-consistency constraints -- could transfer to other inverse problems such as CT reconstruction or super-resolution, provided a robust encoder exists for that modality.","A natural next test is to swap the generation encoder (ViTDet) for another robust feature space (for example, a robust image encoder not trained for detection) to see whether the recall gain is tied to detection-specific features or to semantic alignment in general.","The paper measures recall on the same fastMRI+ annotation distribution used to train the evaluation detector, so the sharpest additional check would be on rare or unseen pathology classes, where the 'small and rare pathologies' motivation is strongest."],"forward_implications":["Applying SDR to any base reconstruction method (diffusion-based or deterministic) raises the pathology recall of the reconstruction set above the single reconstruction and above repeated sampling of the same method.","SDR does not buy recall by adding false positives: on fastMRI+ it also improves mean average precision at IoU 0.25.","Because SDR runs in about three seconds per image, it is fast enough to be used post-hoc in a clinical workflow, unlike diffusion sampling which can take minutes.","The automatic mode (SDR-A) works without any user input, while the manual mode (SDR-M) lets a clinician probe a specific region of concern.","Adversarial robustness of the feature backbone is not optional: using the original non-robust ViTDet does not produce recall or mAP gains over repeated sampling."],"supporting_citations":[{"why":"Establishes that repeated sampling of diffusion models concentrates on likely solutions and yields limited diversity, the problem SDR addresses.","marker":"[5]"},{"why":"Supports the use of features from adversarially robust encoders as a semantic distance metric, the core of the SDR objective.","marker":"[6]"},{"why":"Baseline generative compressed-sensing MRI method (CSGM) that SDR is compared against.","marker":"[11]"},{"why":"Baseline diffusion-based MRI reconstruction method (DDIP) that SDR is compared against.","marker":"[4]"},{"why":"ViTDet object detector whose box features define the semantic distance being maximized.","marker":"[14]"},{"why":"Faster R-CNN detector trained separately to measure pathology recall and mAP in the evaluation.","marker":"[18]"},{"why":"Adversarial fine-tuning technique used to make the ViT backbone robust so that gradients are semantically meaningful.","marker":"[20]"},{"why":"E2E-VarNet baseline reconstruction method that SDR is applied to and compared against.","marker":"[22]"},{"why":"fastMRI+ dataset with knee pathology annotations that provides the training and evaluation data.","marker":"[27]"}],"fun_headline_variants":["SDR: extra MRI reconstructions surface pathologies standard scans miss","Semantically diverse MRI reconstructions reveal rare pathologies overlooked","Multiple MRI reconstructions per scan improve pathology detection","Diverse reconstructions from one MRI scan find more pathologies","SDR: semantically varied MRI images catch false negatives"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation rests on the premise that a Faster-RCNN detector trained on fastMRI+ pathology annotations is a faithful proxy for clinical diagnosis; if detector recall does not track radiologist-visible pathologies, the conclusion that SDR reduces false-negative diagnoses is unsupported even if the measured recall improvement is real.","fun_headline_variants_meta":{"raw":{"variants":["SDR: extra MRI reconstructions surface pathologies standard scans miss","Semantically diverse MRI reconstructions reveal rare pathologies overlooked","Multiple MRI reconstructions per scan improve pathology detection","Diverse reconstructions from one MRI scan find more pathologies","SDR: semantically varied MRI images catch false negatives"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000668,"raw_usage":{"total_tokens":3042,"prompt_tokens":939,"completion_tokens":2103,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":2025}},"tokens_in":555,"tokens_out":2103,"duration_ms":17180,"temperature":1.0,"reasoning_tokens":2025,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:09:59.148724+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Show the fastMRI+ test slices where SDR recovers a pathology that the baseline reconstruction missed to a panel of radiologists, without telling them which reconstruction came from SDR, and ask them to mark the pathologies they can see; if the radiologists do not see the additional pathologies, the claim that SDR uncovers clinically relevant image details is falsified.","supporting_citations":[{"cited_title":"In: ICLR (2024)","cited_arxiv_id":null,"evidence_quote":"Establishes that repeated sampling of diffusion models concentrates on likely solutions and yields limited diversity, the problem SDR addresses."},{"cited_title":"In: IEEE SaTML (2025)","cited_arxiv_id":null,"evidence_quote":"Supports the use of features from adversarially robust encoders as a semantic distance metric, the core of the SDR objective."},{"cited_title":"NeurIPS (2021)","cited_arxiv_id":null,"evidence_quote":"Baseline generative compressed-sensing MRI method (CSGM) that SDR is compared against."},{"cited_title":"In: Leonardis, A., Ricci, E., Roth, S., Russakovsky, O., Sattler, T., Varol, G","cited_arxiv_id":null,"evidence_quote":"Baseline diffusion-based MRI reconstruction method (DDIP) that SDR is compared against."},{"cited_title":"In: ECCV","cited_arxiv_id":null,"evidence_quote":"ViTDet object detector whose box features define the semantic distance being maximized."},{"cited_title":"ICML (2024)","cited_arxiv_id":null,"evidence_quote":"Adversarial fine-tuning technique used to make the ViT backbone robust so that gradients are semantically meaningful."},{"cited_title":"In: Martel, A.L., Abolmaesumi, P., Stoyanov, D., Mateus, D., Zuluaga, M.A., Zhou, S.K., Racoceanu, D., Joskowicz, L","cited_arxiv_id":null,"evidence_quote":"E2E-VarNet baseline reconstruction method that SDR is applied to and compared against."},{"cited_title":"Scientific Data 9(1), 152 (2022)","cited_arxiv_id":null,"evidence_quote":"fastMRI+ dataset with knee pathology annotations that provides the training and evaluation data."}],"review_version":1}