{"id":"b6dc8522-d5fb-4829-b02b-f9071ab775f1","arxiv_id":"1908.07623","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A joint motion estimation and segmentation network, supervised by fully-sampled teacher outputs, predicts cardiac measurements from undersampled MRI nearly as well as from fully-sampled data.","lead":"The paper trains a single neural network to estimate cardiac motion and segment heart structures directly from undersampled MRI data, bypassing the image reconstruction step. A smart generalist should read it because it suggests a path to faster cardiac MRI exams that still produce clinically relevant measurements, though the evidence is currently based on simulated undersampling.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Simulated radial undersampling with synthetic sinusoid phase maps may not transfer to real acquisitions; reported closeness to fully-sampled results could be an artifact of the simulation.","rationale":"The reader's weakest assumption is the same one I would flag: the synthetic sinusoid phase simulation is the foundation of all undersampling artifacts in the experiments. I agree with the conditional verdict because within the simulation the experiments are coherent and support the claim, but external validity is unproven. I do not see an internal inconsistency that would invalidate the simulation results themselves. A secondary gap is that the paper does not compare against a reconstruction-then-analysis pipeline, so the 'bypassing reconstruction' benefit is not directly demonstrated; however, the primary risk remains the realism of the simulated undersampling. The proposed concrete test—evaluating on real complex k-space with retrospectively applied radial masks—would directly test whether the central claim transfers. Therefore the reader's CONDITIONAL verdict is appropriate and should be kept.","tokens_in":6686,"tokens_out":5579,"duration_ms":93235,"concrete_test":"Retrospectively undersample real fully-sampled complex k-space cardiac data (e.g., from CMRxRecon or any dataset with raw k-space and coil sensitivities) using the same radial masks; train and evaluate the proposed model on the resulting images. If Dice/MCD/HD degrade substantially relative to the sinusoid-phase simulation, the simulation assumption fails. Alternatively, replace the sinusoid phase with measured phase maps from complex acquisitions and compare performance across acceleration factors; if sensitivity is high, the reported numbers are not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the equivalence between simulated and real undersampled cardiac MR data. In Section 3 (first paragraph), complex images are generated by adding smoothly varying 2D sinusoid phase maps to magnitude images, then k-space is radially undersampled. This is the only mechanism by which aliasing artifacts are produced. Real cardiac phase maps are not smooth sinusoids: they contain B0 inhomogeneity, flow-induced phase, eddy currents, and tissue-dependent susceptibility. Since the network is trained and evaluated entirely on these synthetic artifacts, the reported robustness to 3x-8x undersampling and the closeness to fully-sampled results may be specific to the smooth-phase simulation rather than to actual accelerated acquisitions. The paper provides no evaluation on retrospectively undersampled real complex k-space data or on phase maps with realistic spatial structure, so the central claim is not yet established for real undersampled cardiac MR.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a joint deep learning model for cardiac motion estimation and segmentation directly from undersampled cardiac MR images, extending the Motion-Seg Net of [6]. The model has motion and segmentation branches sharing a feature encoder, and it is trained with a composite loss that incorporates supervision from fully-sampled image pairs, including a displacement-field matching term (Eq. 2). Experiments on 220 UK Biobank subjects simulate radial undersampling by synthesizing smooth 2D sinusoid phase maps on magnitude images and then back-transforming to k-space. Results in Tables 1 and 2 show that the proposed method outperforms a B-spline FFD baseline and yields Dice and motion errors close to fully-sampled predictions, with the Proposed-B loss improving high-acceleration motion accuracy. The central claim is that clinically relevant motion and segmentation can be obtained directly from undersampled data, bypassing the image reconstruction stage.","tokens_in":6856,"tokens_out":4295,"duration_ms":150057,"significance":"If the results hold, the paper addresses a practically important problem: extracting motion and segmentation from accelerated cardiac MR without an explicit reconstruction step could reduce acquisition time and computational cost. The strengths include the large dataset (220 subjects), evaluation against manual annotations for segmentation accuracy, use of standard metrics (Dice, MCD, HD), and the ablation of the additional displacement-field loss (Proposed-A vs. Proposed-B). The paper also clearly states that at test time only the undersampled network is needed, and it provides a useful comparison with FFD across multiple acceleration factors. However, the significance of the contribution depends on whether the simulated undersampling reflects real accelerated acquisitions and on whether direct estimation is compared with the conventional reconstruct-then-analyze pipeline.","major_comments":[{"comment":"The evaluation is entirely based on synthetic complex-valued images generated by adding smooth 2D sinusoid phase maps to magnitude images (following [10]), and all undersampling artifacts in both training and testing arise from this simulation. Real cardiac phase maps contain B0 inhomogeneity, flow-induced phase, eddy currents, and tissue-dependent susceptibility, which are not smooth sinusoids. Therefore the central claim that the model is robust to undersampled data and produces results close to fully-sampled ones is not yet established for real accelerated cardiac MR. I request additional experiments on retrospectively undersampled real complex k-space data, or, if such data are unavailable, a clear statement of this limitation and a discussion of how the smooth-phase simulation might affect transferability.","section":"Section 3, paragraph 1"},{"comment":"The claim of bypassing the reconstruction stage is not supported by a reconstruction-then-analysis comparison. The paper compares only against FFD [8] and the fully-sampled joint model [6]; a natural and important baseline is to reconstruct undersampled k-space with a standard method and then run the fully-sampled Motion-Seg Net on the reconstructed images, reporting the same metrics (MCD, HD, Dice, EF). Without such a baseline, the reader cannot judge whether direct joint prediction is preferable to the conventional pipeline. In addition, the segmentation results are not compared with the closest prior work [10], which predicts segmentation maps directly from undersampled k-space; this comparison is needed to situate the segmentation contribution.","section":"Section 3, Tables 1 and 2"},{"comment":"The motion supervision in Eq. (2) uses displacement fields ∆FS predicted by the fully-sampled sub-network, which is the authors' own network from [6]. Because ∆FS is not an independent ground truth, biases in the fully-sampled network are inherited by the undersampled network, and the statement that 'no ground truth displacement fields are required' is potentially misleading: this is a form of knowledge distillation rather than purely unsupervised learning. The ROI mask M_t is also generated from the fully-sampled segmentation predictions, so the supervision itself is not independent of the teacher, even though the evaluation metrics are computed against manual annotations. Please clarify the role of the teacher, discuss the potential effect of teacher bias, and consider whether this weakens the 'unsupervised' claim.","section":"Section 2.1, Eq. (2)"},{"comment":"The claim that undersampled results are 'close to' fully-sampled results is not supported by statistical tests. For example, Table 1 shows Proposed-B at 8x with LV MCD 1.79 mm versus 1.30 mm for the fully-sampled Joint Model [6], and Table 2 shows Myo Dice dropping from 0.8640 (FS) to 0.8260 at 8x. These differences may be clinically meaningful. Please report whether the differences between undersampled and fully-sampled results are statistically significant, and discuss their clinical relevance in the context of cardiac function assessment.","section":"Section 3, Tables 1 and 2"}],"minor_comments":[{"comment":"The EF error is computed relative to fully-sampled predictions rather than to manual ground truth; consider also reporting EF accuracy against manual segmentations, since the fully-sampled predictions themselves contain errors.","section":"Table 2"},{"comment":"The description of the data split (100/100/20) should explicitly state that the split is at the subject level and that no slices from the same subject appear in both training and test sets.","section":"Section 3, paragraph 2"},{"comment":"The warp operator Γ is used in the equations but is not defined; please provide a definition or a reference for the bilinear interpolation warping operation.","section":"Equations (1) and (2)"},{"comment":"The text mentions that the loss Lw is not employed, but Lw is never defined; please define it or rephrase the sentence to be self-contained.","section":"Section 2.2"},{"comment":"References [6] and [7] are arXiv preprints; please update them to the peer-reviewed versions if they are now published.","section":"References"},{"comment":"The labels '8 × 1 15 20 25' in Fig. 2 are unclear; please clarify which numbers denote acceleration factor and which denote time points.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The simulation-realism concern is the main issue: all experiments use synthetic phase maps, so the central claim is only established for the simulation, not for real undersampled acquisitions. The missing reconstruction-then-analysis baseline is equally important, as it directly tests the 'bypassing reconstruction' premise. The use of the authors' own fully-sampled network as a teacher is acceptable in principle but should be disclosed clearly and its effect on the motion evaluation discussed. These issues are addressable with additional experiments or a careful limitation section, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does one useful thing: it takes the joint motion-segmentation network from [6] and adapts it to work directly on undersampled k-space data by adding a fully-sampled teacher branch and a displacement-field loss. That is a modest but legitimate extension, and the training scheme is reasonable: the teacher provides supervision for motion and segmentation, while the final evaluation is against manual annotations, so the self-referential teacher does not contaminate the metrics. On 220 UK Biobank subjects, the method holds up well at 3x–8x radial undersampling, with small drops in Dice and motion errors, and the Proposed-B loss clearly helps at higher acceleration factors. The paper is clearly written and the authors are honest that this is simulation-based.\n\nThe soft spots are real but not fatal. The stress-test note is on target: the undersampled data are generated by adding smooth sinusoid phase maps to magnitude images, then retrospectively undersampling k-space. Real cardiac phase maps contain flow-induced phase, B0 inhomogeneity, and eddy currents, so the network may be learning to handle an unrealistically smooth artifact. This is the main reason the central claim is not yet established for clinical use. The paper also never compares against Schlemper et al. [10], the closest segmentation-from-undersampled baseline, nor against a reconstruction-then-analysis pipeline, so it is hard to say what is actually gained by bypassing reconstruction. There is no code or data release, and hyperparameter choices (alpha, beta, lambda) are reported but not analyzed beyond a validation-set selection.\n\nThe citation pattern is fine; the authors rely heavily on their own prior work, but that is legitimate since the method is a direct extension. The statistical testing (Wilcoxon) is a plus. Overall, this is a competent conference paper with a clear contribution to the application-driven MRI line of work. It deserves a serious peer review, but a reviewer should push for evaluation on retrospectively undersampled real complex k-space data (which the UK Biobank complex data would allow) or at least on more realistic simulated artifacts, before accepting the claim that the model is robust to actual accelerated acquisitions.","headline":"A sensible engineering extension of the authors' own joint motion-segmentation model to undersampled cardiac MR, but the simulation-only evaluation leaves the main clinical claim unproven.","tokens_in":659,"tokens_out":1775,"would_cite":false,"duration_ms":62283,"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 joint deep network can estimate cardiac motion and segmentation directly from undersampled MR data, bypassing image reconstruction, with accuracy close to fully-sampled predictions.","keywords":["cardiac MRI","motion estimation","image segmentation","undersampled k-space","deep learning","joint learning","ejection fraction","bypassing reconstruction"],"falsifier":"Acquire prospectively undersampled cardiac MR images with fully sampled references for the same subjects and compare the network's motion and segmentation accuracy, trained only on the synthetic phase-map simulation, against the fully sampled predictions. If the accuracy gap at the same acceleration factors is materially larger than the 3x, 6x, and 8x gaps reported in the paper, the central claim is not supported.","tokens_in":6526,"feed_emoji":"🫀","tokens_out":7631,"duration_ms":232323,"temperature":0.7,"pith_summary":"This paper tries to establish that cardiac motion fields and segmentation maps can be predicted directly from undersampled MR data, without first reconstructing the images. The motivation is that clinical measurements such as ejection fraction may stay reliable even when the images themselves are imperfect, so reconstruction can be skipped when the goal is quantitative analysis. The authors build a joint network with a motion-estimation branch and a segmentation branch, and train it with a parallel fully-sampled sub-network that guides the undersampled branch. On 220 cardiac MR sequences at 3x, 6x, and 8x radial undersampling, the reported motion and segmentation accuracy is close to the fully-sampled baseline, with ejection fraction errors of 2.68%, 3.56%, and 4.16%. If this holds, faster cardiac scans could feed directly into motion and contour analysis for diagnosis.","feed_headline":"Heart motion and contours from undersampled MRI, no reconstruction","feed_subtitle":"Trained with fully sampled guidance, one network handles both tasks from undersampled scans.","key_machinery":"The central object is the joint Motion-Seg network, a shared feature encoder with a motion-estimation branch and a segmentation branch, extended for undersampled inputs by a parallel fully-sampled sub-network that operates only during training. The motion branch estimates a pixel-wise 2D displacement field between two cardiac frames and uses it to warp the fully-sampled source image; the segmentation branch outputs label maps for the left-ventricular cavity, myocardium, and right-ventricular cavity. The load-bearing supervisory mechanism is the loss in Eq. (2), which combines the intensity error between the warped fully-sampled source and target, a gradient-based smoothness term on the displacement field, and a masked mean-squared error between the undersampled and fully-sampled displacement fields restricted to the region of interest. This lets the undersampled network learn from the fully-sampled network without any ground-truth motion fields.","core_discovery":"The paper's central claim is that one shared deep network can estimate cardiac motion and segmentation jointly from undersampled k-space data and produce results close to those obtained from fully-sampled data. During training, a pre-trained sub-network on corresponding fully-sampled image pairs supervises the undersampled path: the undersampled motion field warps the fully-sampled source image, and the difference between undersampled and fully-sampled displacement fields is penalized inside a region of interest derived from segmentation. Segmentation supervision combines manual labels where available with predictions from the fully-sampled sub-network elsewhere. At test time only the undersampled branch is used. Reported contour distances and Dice scores degrade mildly from 3x to 8x acceleration, and the clinical parameter ejection fraction stays within about four percentage points of the fully-sampled value.","pith_inferences":["If the synthetic phase-map simulation is a faithful proxy for real cardiac phase, the same guided-training scheme could transfer to other dynamic imaging settings where a fully-sampled reference is available during training.","A natural extension is to re-enable the warped-segmentation consistency loss with uncertainty weighting for low acceleration factors, since the paper found it harmful at high undersampling.","The reported ceiling is set by the fully-sampled network's own accuracy; at very high acceleration, the displacement-field supervision may need to be replaced by a learned prior independent of the fully-sampled reference.","The current evidence is retrospective and simulated; a prospective study with real accelerated acquisitions would test whether the training transfer holds."],"forward_implications":["At 8x acceleration, left-ventricular Dice is 0.9141 versus 0.9348 on fully-sampled data, and ejection fraction error is 4.16%, so for these clinical measures the reconstruction stage may be safely omitted.","The displacement-field supervision term (Proposed-B) mainly improves motion accuracy at 6x and 8x, indicating that stronger guidance is needed as aliasing increases.","Because the fully-sampled sub-network is only needed at training time, the same model can be deployed with undersampled acquisitions alone.","The paper's decision to drop the warped-segmentation consistency loss suggests that cross-task consistency constraints can add noise under undersampling, and should be re-examined when designing joint models for accelerated data."],"supporting_citations":[{"why":"Supplies the baseline Motion-Seg network architecture and the joint fully-sampled loss that this paper extends to undersampled inputs.","marker":"[6]"},{"why":"Provides the phase-map synthesis scheme used to turn magnitude images into complex images from which undersampled k-space data are simulated.","marker":"[10]"},{"why":"Provides the free-form deformation registration algorithm used as the motion-estimation baseline and the multi-level registration idea.","marker":"[8]"},{"why":"Supplies the fully convolutional segmentation architecture used for the segmentation branch.","marker":"[1]"},{"why":"Establishes the application-driven MRI idea of predicting clinical outputs directly from undersampled data.","marker":"[2]"},{"why":"Supplies the Huber-like gradient loss approximation used for displacement-field smoothness regularization.","marker":"[3]"}],"fun_headline_variants":["Skip reconstruction: joint motion and contours from undersampled MRI","One network: motion + segmentation straight from undersampled cardiac MRI","Undersampled cardiac MRI goes straight to motion and segmentation","Direct motion and segmentation from undersampled cardiac scans"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training pipeline assumes that undersampled k-space data can be realistically simulated by adding smooth 2D sinusoid phase maps to magnitude images; if real cardiac phase varies more rapidly or structurally, a network trained on these simulations may not transfer to actual accelerated acquisitions.","fun_headline_variants_meta":{"raw":{"variants":["Skip reconstruction: joint motion and contours from undersampled MRI","One network: motion + segmentation straight from undersampled cardiac MRI","Undersampled cardiac MRI goes straight to motion and segmentation","Direct motion and segmentation from undersampled cardiac scans"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000377,"raw_usage":{"total_tokens":1982,"prompt_tokens":892,"completion_tokens":1090,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":1020}},"tokens_in":508,"tokens_out":1090,"duration_ms":8540,"temperature":1.0,"reasoning_tokens":1020,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:00:24.277240+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Acquire prospectively undersampled cardiac MR images with fully sampled references for the same subjects and compare the network's motion and segmentation accuracy, trained only on the synthetic phase-map simulation, against the fully sampled predictions. If the accuracy gap at the same acceleration factors is materially larger than the 3x, 6x, and 8x gaps reported in the paper, the central claim is not supported.","supporting_citations":[{"cited_title":"Joint Learning of Motion Estimation and Segmentation for Cardiac MR Image Sequences","cited_arxiv_id":"1806.04066","evidence_quote":"Supplies the baseline Motion-Seg network architecture and the joint fully-sampled loss that this paper extends to undersampled inputs."},{"cited_title":"In: MICCAI (2018)","cited_arxiv_id":null,"evidence_quote":"Provides the phase-map synthesis scheme used to turn magnitude images into complex images from which undersampled k-space data are simulated."},{"cited_title":"IEEE transactions on medical imaging 18(8), 712–721 (1999)","cited_arxiv_id":null,"evidence_quote":"Provides the free-form deformation registration algorithm used as the motion-estimation baseline and the multi-level registration idea."},{"cited_title":"Journal of Cardiovascular Magnetic Resonance (2018)","cited_arxiv_id":null,"evidence_quote":"Supplies the fully convolutional segmentation architecture used for the segmentation branch."},{"cited_title":"In: International Con- ference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"Establishes the application-driven MRI idea of predicting clinical outputs directly from undersampled data."},{"cited_title":"In: CVPR (2017)","cited_arxiv_id":null,"evidence_quote":"Supplies the Huber-like gradient loss approximation used for displacement-field smoothness regularization."}],"review_version":1}