{"id":"aecfd35e-2d0b-4ced-86a3-0c6d8a6ecb17","arxiv_id":"2506.19167","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"FLIR is a cascade of lightweight U-Nets for cardiac MRI registration; it beats VoxelMorph dice scores only with 3 to 5 cascades, which run slower than VoxelMorph, so the speed claim is not supported.","lead":"This thesis presents FLIR, a cascaded U-Net model that registers cardiac MRI volumes to measure heart strain. It claims similar registration accuracy to published state-of-the-art methods with less added inference time, and consistent strain values on test-retest scans.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The speed-fidelity claim is contradicted by the thesis's own tables: 1xFLIR is faster but less accurate than VoxelMorph, while every FLIR variant that matches or beats VoxelMorph's dice is slower; the RC-Net/VR-Net timing comparison uses cross-dataset numbers with a seconds-vs-ms unit mismatch.","rationale":"The reader's reject verdict is well supported. I part company slightly on which assumption is weakest: the affine-motion assumption is a domain-scope caveat that the authors explicitly acknowledge in Section 5.2, and it does not by itself invalidate the reported numbers. The more fundamental problem is internal to the evidence: the FLIR family never dominates VoxelMorph in both speed and accuracy. Since the thesis's motivating contribution is a speed advantage with comparable fidelity, this is the load-bearing claim. The paper deserves credit for reporting full timing tables and a unit-tested strain calculator (Appendix A), and the longitudinal strain repeatability (0.032 average difference) is a real positive result; but radial and circumferential repeatability are roughly ten times worse and are presented without a VoxelMorph or SOTA baseline, so they cannot serve as independent support. The proposed controlled re-measurement would settle whether any FLIR point is Pareto-superior to VoxelMorph and whether the RC-Net/VR-Net overhead comparison survives unit correction.","tokens_in":47390,"tokens_out":5445,"duration_ms":56568,"concrete_test":"Re-run a controlled comparison on one shared test bed: train VoxelMorph, RC-Net (2x/3x), VR-Net, and FLIR (1x/2x/3x/5x) on the same ACDC+UKBB training split; measure LV/Myo/RV dice on the same test set and wall-clock inference with the same PyTorch profiler on the same CPU and GPU, using one time unit. Then check whether any FLIR cascade satisfies dice >= VoxelMorph and time < VoxelMorph. If no such point exists, or if in-house RC-Net/VR-Net outperform FLIR at matched runtime, the central speed-fidelity claim is falsified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The contribution's core claim is that FLIR achieves state-of-the-art fidelity at a fraction of the inference time. The thesis's own measurements do not contain any operating point with that property. In Table 4.1, 1xFLIR has LV dice 0.801, below VoxelMorph's 0.835; in Table 4.2, 1xFLIR is indeed faster (47.2 ms vs 76.2 ms CPU). The next cascade, 2xFLIR, just matches VoxelMorph (0.839 vs 0.835) but already takes 98.9 ms, 1.3x longer; 3xFLIR and 5xFLIR improve dice (0.864, 0.882) but take 1.7x and 3.1x VoxelMorph's CPU time. Thus the abstract's speed claim only holds for a variant that loses fidelity, and the fidelity claim only holds for variants that are slower than the baseline. The comparison to RC-Net and VR-Net does not repair this: Tables 4.4-4.6 import Jia et al.'s numbers from different datasets and hardware, and Table 4.6 labels values as milliseconds although those same values (5.97, 11.95, ...) are seconds in Table 2.3. The 'smaller overhead than RC-Net/VR-Net' conclusion therefore rests on a cross-dataset, unit-inconsistent comparison rather than a controlled speed-fidelity Pareto measurement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript (an MSc thesis posted on arXiv) proposes FLIR, a cascaded U-Net for unsupervised 3D registration of cardiac cine SSFP images at end-diastole and end-systole, together with a strain calculator that converts the predicted deformation into radial, circumferential, and longitudinal strain. The method is evaluated against a VoxelMorph baseline trained on the same ACDC/UKBB data, compared with published RC-Net, VR-Net, and VoxelMorph numbers from Jia et al., and tested for strain repeatability on a 10-patient test-retest dataset. The central claim is that FLIR achieves registration fidelity comparable to state-of-the-art methods while taking a fraction of the inference time. The paper's own measurements, however, do not support this claim: the only FLIR variant faster than VoxelMorph is 1xFLIR, which has lower Dice, while all variants that match or exceed VoxelMorph Dice are slower. The cross-paper speed comparison also contains a units error and uses non-controlled conditions.","tokens_in":47741,"tokens_out":8251,"duration_ms":81793,"significance":"If the speed-fidelity claim were established, the work would be practically useful: a lightweight cascaded registration model with a documented strain calculator and test-retest strain repeatability could lower the computational barrier to large-scale CMR strain analysis. The paper has some genuine strengths: the VoxelMorph-Lite ablation cleanly isolates the runtime cost of full-resolution layers, the strain-calculator unit tests in Appendix A are a reproducible verification artifact, and the limitation discussion in Section 5.2 is candid. However, the central contribution—a model that is both faster than VoxelMorph and at least as accurate—is not present in the paper's own tables, and the external speed comparison relies on cross-dataset numbers with an order-of-magnitude unit error. As presented, the main claim is not established, although the underlying engineering (efficient cascaded U-Nets for cardiac registration) is a reasonable starting point for further work.","major_comments":[{"comment":"The central claim that FLIR performs registration in a fraction of the time while maintaining fidelity is contradicted by the paper's own measurements. In Table 4.1, 1xFLIR has lower LV Dice than VoxelMorph (0.801 vs 0.835) and is the only FLIR variant that is faster on CPU (47.2 ms vs 76.2 ms in Table 4.2). Every configuration that matches or exceeds VoxelMorph LV Dice—2x, 3x, and 5x FLIR—is slower on CPU (98.9, 133.1, and 238.0 ms vs 76.2 ms), with the same ordering on GPU1 and GPU2. Thus the results contain no operating point that is simultaneously faster than VoxelMorph and at least as accurate; the abstract and the Section 5.1 conclusion overstate what Table 4.2 supports.","section":"§4.2.2, Tables 4.1 and 4.2"},{"comment":"The speed comparison against RC-Net and VR-Net is invalid as reported. Table 4.6 labels the values from Jia et al. as milliseconds, but Table 2.3, the source of those exact numbers (5.97, 0.10, 11.95, 0.21, etc.), reports seconds. The table therefore misstates the external runtimes by three orders of magnitude. Additionally, the RC-Net/VR-Net Dice and timing figures in Tables 4.4–4.6 come from different datasets (UKBB, 3DCMR), different preprocessing, and different hardware than the FLIR evaluation, so the relative-slowdown comparisons in Section 4.2.3 and Figures 4.8–4.11 are not controlled comparisons. The conclusion that FLIR has smaller overhead than RC-Net/VR-Net rests on these inconsistent numbers.","section":"§4.2.2, Table 4.6 and Table 2.3"},{"comment":"The decision to remove all affine registration stages is load-bearing for the method's speed, but the underlying assumption is untested. The thesis assumes that for any clinically usable SSFP sequence there should be no affine deformation (§3.1), yet ED and ES volumes from the same study can still be affected by breath-hold position changes, through-plane motion, and slice misregistration. Section 5.2 explicitly limits the evaluation to artifact-free data. Without an experiment or analysis showing that residual affine components are negligible for the target data, the claimed speed advantage may not transfer to realistic acquisitions, because the deformable network would have to compensate for global transformations itself.","section":"§3.1 and §5.2"},{"comment":"The test-retest strain analysis contains an unexplained duplication and a metrics problem. In Table 4.7, Patient 10 has identical test and retest values for all three strain components (0.763, 0.932, 0.818), giving zero differences in Table 4.8; for a genuine retest acquisition this is implausible and is not commented on. More broadly, the average test-retest differences reported in Table 4.9 for radial (0.332±0.236) and circumferential (0.344±0.356) strain are large relative to the myocardial strain values the method is intended to measure, and the authors show in Section 4.3.2 that the nominally peak strains are dominated by blood-pool artifacts unless the most negative values are selected (Tables 4.10–4.11). The claim that strain values are very consistent is therefore only supported for longitudinal strain and only after a post-hoc redefinition of the peak metric.","section":"§4.3.2, Tables 4.7–4.11"}],"minor_comments":[{"comment":"The compiled GPU1 total timings (e.g., 0.0269 ms for VoxelMorph) are about two orders of magnitude smaller than the uncompiled values in Table 4.2 (2.193 ms for the same configuration); the measurement protocol, including whether data transfer and warping are included, should be stated.","section":"Table 4.3"},{"comment":"In Table 4.10, the Patient 3 retest value 0.441 is positive under a column labeled 'Peak Negative Radial Strain'; if this is a sign typo, it should be corrected.","section":"Table 4.10"},{"comment":"The strain tables do not state the units of the reported strain values or define whether 'peak' means maximum absolute value, maximum positive value, or most negative value; this makes the values in Table 4.7 difficult to interpret.","section":"Tables 4.7–4.11"},{"comment":"Several typographical errors remain, including 'Convolutoinal' in the list of abbreviations, 'start-of-the-art' in Section 2.7, and 'inevitability loss' in Section 3.1 (presumably 'invertibility loss').","section":"Throughout"},{"comment":"The abbreviation LDDMM is used in the discussion of Quicksilver without definition or citation; a brief definition and reference would help readers outside the registration subfield.","section":"§2.5.1"}],"recommendation":"reject","confidential_remarks":"This is a thesis-style manuscript with an extensive tutorial component. The core empirical claim is contradicted by the paper's own tables, and the external speed comparison contains a unit error and non-controlled conditions. I do not see a path to acceptance without either new experiments providing a faster-and-accurate operating point or a substantial reframing of the contribution, which is beyond a minor revision. The duplicate Patient 10 strain values and the Table 4.6 unit error also suggest that the experimental records should be carefully re-checked before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline claim doesn't survive contact with the thesis's own tables. Across the measured operating points, no FLIR variant is both faster than VoxelMorph and as accurate: 1xFLIR is faster (47 ms vs 76 ms CPU) but has lower LV dice (0.801 vs 0.835); 2xFLIR matches dice but takes 1.3x longer; 3x and 5x are more accurate but 1.7x and 3.1x slower. So the abstract's 'fraction of the time' only holds for the least accurate variant, and the fidelity gain comes only with a slowdown. The comparison to RC-Net and VR-Net also leans on a units error: Table 4.6 labels values as milliseconds that are seconds in Table 2.3, and the numbers come from a different dataset and hardware. That is load-bearing because it is the basis for the claimed efficiency advantage over state of the art.\n\nThe thesis does several things well. The VoxelMorph-Lite experiment cleanly shows the cost of full-resolution layers, and the discussion of cascade trade-offs is sensible. The strain calculator comes with unit tests, which is more than most papers do. The limitations section is candid about the no-affine assumption, blood-pool contamination, and the cylinder-center assumption. The architecture is described well enough to reproduce, and the within-dataset dice comparisons are fairly presented.\n\nThe main soft spots beyond the units issue: the strain repeatability experiment has no baseline (no VoxelMorph), so the consistency numbers are hard to interpret. The test-retest dataset is only 10 patients, and the radial/circumferential differences (0.33-0.34) are an order of magnitude larger than longitudinal (0.032), which is not adequately discussed. The paper would be substantially improved by reframing the contribution as a favorable dice-per-time tradeoff rather than an absolute speed advantage, and by correcting the unit error.\n\nThis is a Master's thesis, not a major advance, but the empirical work is real and the topic is relevant. It deserves a serious referee because the flaws are identifiable and fixable; the current version should not be accepted without major revision.","headline":"A clearly written thesis whose headline speed claim is contradicted by its own tables; the efficiency analysis has value but the central comparison is broken by a units error.","tokens_in":48250,"tokens_out":3478,"would_cite":false,"duration_ms":36127,"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":"FLIR, a cascade of lightweight U-Nets, registers cardiac MR volumes with fidelity comparable to state-of-the-art methods at a fraction of their inference cost, with strain values that repeat between test and retest acquisitions.","keywords":["cardiac magnetic resonance imaging","image registration","deep learning","cascaded U-Net","cine SSFP","strain quantification","inference speed","test-retest repeatability"],"falsifier":"Register a batch of cardiac volume pairs to which small synthetic rigid offsets have been applied (for example 2-5 mm translations or 1-2 degree rotations between the fixed and moving volumes), then measure whether FLIR's LV dice and test-retest strain agreement fall below VoxelMorph's; a drop would confirm that the no-affine assumption, not the cascade design, is carrying part of the result.","tokens_in":47184,"feed_emoji":"🫀","tokens_out":11258,"duration_ms":103398,"temperature":0.7,"pith_summary":"The paper sets out to show that cardiac magnetic resonance (CMR) image registration — the step that turns a beating-heart cine movie into a map of tissue motion — does not have to choose between accuracy and speed. It proposes FLIR, a cascade of lightweight U-Nets, and claims that this design registers end-diastolic to end-systolic volumes from the same cine SSFP acquisition with fidelity comparable to state-of-the-art registration networks while running at a fraction of their inference cost. The speed comes from two design decisions: dropping the affine-alignment stage that other architectures treat as standard, and concentrating convolution operations on downsampled volumes where they are cheap. The paper then argues that the payoff is practical: the predicted deformations feed a finite-strain calculator, and the resulting radial, circumferential, and longitudinal strain values agree closely between test and retest acquisitions of the same patient — the consistency a motion-based alternative to gadolinium-enhanced imaging would need. If the claims hold, cardiac strain quantification from a routine contrast-free SSFP scan becomes a realistic research and clinical option.","feed_headline":"Fast U-Net cascade tops VoxelMorph on cardiac MR registration","feed_subtitle":"More registration fidelity per millisecond than bulkier rivals, with strain reproducible across repeated scans.","key_machinery":"The load-bearing mechanism is a cascaded stack of compact U-Nets, each built so that convolutions run mostly at low resolution: the first layers use a 3x3x3 kernel with stride 2, and channel counts double as resolution halves, keeping full-resolution operations to a minimum. Each cascade outputs a displacement field $\\phi$; the warped moving image $I_W$ is scored against the fixed image $I_F$ with the correlation-coefficient loss $L_{\\text{corrcoef}} = 1 - \\text{CorrCoef}[I_W, I_F]$, plus a total-variation smoothness term $L_{TV}$ that penalizes discontinuities in the field. The per-cascade similarity loss is back-propagated to every previous cascade, following the VTN training format. Two domain assumptions carry the design: same-acquisition SSFP volumes need no affine registration because patient motion would corrupt the images, and the left ventricle can be treated as a cylinder centered in the volume when converting Cartesian strain to radial, circumferential, and longitudinal strain. The companion strain calculator converts the flow field into finite strain via the deformation gradient and the right Cauchy-Green tensor, in both Lagrangian and Eulerian frames.","core_discovery":"The central claim is that a single-pass, volumetric deep learning registration model can match the fidelity of state-of-the-art cardiac CMR registration without paying its inference cost, provided the architecture is engineered around a domain assumption: the two volumes come from one clinically usable SSFP acquisition, so no affine alignment is needed. FLIR is a cascade of U-Net sub-networks whose first convolutions use a stride of 2, so most operations run on downsampled volumes; each cascade predicts a small displacement field, the warped image is scored against the fixed image with the correlation-coefficient loss, and a total-variation term keeps the fields smooth. On the author's evaluation, 5xFLIR reaches a left-ventricle dice of 0.882 versus 0.835 for VoxelMorph on the same data, the largest fidelity gain of the compared methods, while 3xFLIR runs at about 1.7x VoxelMorph's CPU time compared with roughly 2.0x and 3.1x for RC-Net 2x and VR-Net relative to their own baselines. The same flow fields produce repeatable strain estimates, with average test-retest differences of 0.332 for radial, 0.344 for circumferential, and 0.032 for longitudinal strain, and the strain maps flag abnormal longitudinal strain in patients whose tissue is unhealthy, consistent with the physiology the thesis relies on. The stated goal of the thesis is that such a fast, optimized pipeline makes strain quantification from SSFP images a credible path to inferring unhealthy tissue without gadolinium contrast.","pith_inferences":["The same architecture should transfer to other intra-acquisition cardiac time-point pairs, not only end-diastole to end-systole, because neither design decision depends on which two phases are being registered.","The strain repeatability result is the strongest evidence for clinical utility, but the thesis does not compare its strain values against tagging-based or expert-measured ground truth; such a comparison would be the natural validation next step.","The paper itself notes that peak strain readings are contaminated by large values in the blood pool; masking the myocardium with a segmentation before computing strain would likely tighten the radial and circumferential test-retest differences further.","Other volumetric registration problems whose inputs come from a single acquisition — for example some fetal or fast abdominal sequences — could adopt the same 'drop the affine stage' shortcut and capture a similar speedup."],"forward_implications":["Fidelity scales with cascade count: one cascade runs faster than VoxelMorph but scores lower, while three and five cascades surpass VoxelMorph on left-ventricle, myocardium, and right-ventricle dice.","The 3-cascade model reaches fidelity close to RC-Net and VR-Net at about 1.7x VoxelMorph's CPU time, versus roughly 2.0x and 3.1x for those competitors, making the accuracy-per-runtime tradeoff the model's central advantage.","The deformations are stable enough for strain quantification: average test-retest differences are 0.332 (radial), 0.344 (circumferential), and 0.032 (longitudinal), so longitudinal strain in particular repeats tightly.","The no-affine assumption means same-acquisition cardiac registration does not need alignment cascades, which is what lets the whole pipeline stay light.","Fast inference on mid-level hardware makes cohort-scale processing feasible, which the thesis cites as a precondition for research and clinical adoption."],"supporting_citations":[{"why":"Supplies the VoxelMorph baseline network and the U-Net-style registration template that FLIR modifies, along with the observation that full-resolution layers dominate runtime.","marker":"[7]"},{"why":"VTN's cascading architecture with per-cascade similarity loss, which FLIR adopts as its training format.","marker":"[73]"},{"why":"RC-Net, the cascading competitor whose fidelity and slowdown figures anchor the claim that FLIR achieves comparable dice at lower relative cost.","marker":"[72]"},{"why":"Provides VR-Net and the comparative performance tables that give FLIR its state-of-the-art dice and inference-time benchmarks.","marker":"[30]"},{"why":"The U-Net architecture used as the template for each FLIR sub-network.","marker":"[54]"},{"why":"The public ACDC dataset that supplies training, validation, and test volumes for the registration experiments.","marker":"[11]"},{"why":"The UK Biobank dataset used alongside ACDC for training and evaluation, whose scale motivates the need for fast inference.","marker":"[57]"},{"why":"Defines the dice score, the similarity metric used to evaluate registration fidelity throughout the experiments.","marker":"[12]"},{"why":"Documents how patient motion corrupts MRI acquisitions, forming the basis for assuming no affine deformation between same-acquisition volumes.","marker":"[70]"}],"fun_headline_variants":["Fast U-Net cascade matches state-of-the-art cardiac MR registration","Cardiac MR registration: speed without sacrificing fidelity","FLIR: fast cardiac strain mapping with repeatable outputs","Speedy cardiac MR registration with consistent strain estimates"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that any two volumes registered by FLIR come from the same clinically usable SSFP acquisition, so that no affine alignment is needed; if patient motion, breath-hold shifts, or slice misregistration actually separate the volumes, the deformable field alone must absorb them and the reported speed-fidelity tradeoff would not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Fast U-Net cascade matches state-of-the-art cardiac MR registration","Cardiac MR registration: speed without sacrificing fidelity","FLIR: fast cardiac strain mapping with repeatable outputs","Speedy cardiac MR registration with consistent strain estimates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001196,"raw_usage":{"total_tokens":5040,"prompt_tokens":1158,"completion_tokens":3882,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":774,"completion_tokens_details":{"reasoning_tokens":3817}},"tokens_in":774,"tokens_out":3882,"duration_ms":30522,"temperature":1.0,"reasoning_tokens":3817,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:07:20.320718+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Register a batch of cardiac volume pairs to which small synthetic rigid offsets have been applied (for example 2-5 mm translations or 1-2 degree rotations between the fixed and moving volumes), then measure whether FLIR's LV dice and test-retest strain agreement fall below VoxelMorph's; a drop would confirm that the no-affine assumption, not the cascade design, is carrying part of the result.","supporting_citations":[{"cited_title":"Voxelmorph: a learning framework for deformable medical image registration.IEEE transactions on medical imaging, 38(8):1788–1800, 2019","cited_arxiv_id":null,"evidence_quote":"Supplies the VoxelMorph baseline network and the U-Net-style registration template that FLIR modifies, along with the observation that full-resolution layers dominate runtime."},{"cited_title":"Unsupervised 3d end-to- end medical image registration with volume tweening network.IEEE journal of biomedical and health informatics, 24(5):1394–1404, 2019","cited_arxiv_id":null,"evidence_quote":"VTN's cascading architecture with per-cascade similarity loss, which FLIR adopts as its training format."},{"cited_title":"Recursive cascaded networks for unsupervised medical image registration","cited_arxiv_id":null,"evidence_quote":"RC-Net, the cascading competitor whose fidelity and slowdown figures anchor the claim that FLIR achieves comparable dice at lower relative cost."},{"cited_title":"Learning a model-driven variational network for deformable image registration.IEEE Transactions on Medical Imaging, 41(1):199–212, 2021","cited_arxiv_id":null,"evidence_quote":"Provides VR-Net and the comparative performance tables that give FLIR its state-of-the-art dice and inference-time benchmarks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The public ACDC dataset that supplies training, validation, and test volumes for the registration experiments."},{"cited_title":"Uk biobank: an open access resource for identifying the causes of a wide range of complex diseases of middle and old age.PLoS medicine, 12(3):e1001779, 2015","cited_arxiv_id":null,"evidence_quote":"The UK Biobank dataset used alongside ACDC for training and evaluation, whose scale motivates the need for fast inference."},{"cited_title":"Optimizing the dice score and jaccard index for medical image segmentation: Theory and practice","cited_arxiv_id":null,"evidence_quote":"Defines the dice score, the similarity metric used to evaluate registration fidelity throughout the experiments."},{"cited_title":"Motion artifacts in mri: A complex problem with many partial solutions.Journal of Magnetic Resonance Imaging, 42(4):887–901, 2015","cited_arxiv_id":null,"evidence_quote":"Documents how patient motion corrupts MRI acquisitions, forming the basis for assuming no affine deformation between same-acquisition volumes."}],"review_version":1}