{"id":"fceb1e7e-3463-4e62-9fb1-2d54811d9d93","arxiv_id":"1908.09140","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"LANTERN unrolls an ADMM compressed-sensing model into a convolutional network with DCT and total-variation initialization, and reports higher PSNR/SSIM than k-t SLR and D5C5 for dynamic cardiac MRI using only 100 training volumes.","lead":"This paper presents LANTERN, a deep learning method that reconstructs dynamic cardiac MRI from undersampled data and claims to work well with only 100 training scans. It matters because deep-learning MRI methods typically need large datasets, and the small-data regime is the practical bottleneck in clinical imaging.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central small-data claim is not established: D5C5 is never trained on the same 100 volumes, and the paper explicitly states this in the Implementation section.","rationale":"The reader's weakest_assumption identifies exactly the same gap: D5C5 is trained on 2900 volumes versus 100 for LANTERN, with no same-size control. This is the single most load-bearing concern because the abstract, introduction, and discussion all frame the contribution as superior accuracy with limited data. If D5C5 also reconstructs well when trained on 100 volumes, the comparison in Tables 4 and 5 does not establish the claimed advantage; if D5C5 performs poorly at 100 volumes, the claim still needs that experiment to be shown. The paper's own Implementation paragraph flags this limitation, and the Impact of dataset size section compounds it by selecting the training-set size using the test set. I do not see an internal inconsistency in the network derivation or the reported metrics, and the DCT+TV initialization and ADMM unrolling are reasonable design choices with independent motivation from earlier unrolled networks. However, the central empirical claim is not cleanly supported without the same-data-size comparison, so the verdict should remain conditional rather than accept or reject outright. No code or data are released, which makes the missing control harder to verify, but the request for a same-size D5C5 experiment and a proper validation split is the concrete path to resolving the concern.","tokens_in":14318,"tokens_out":3327,"duration_ms":36150,"concrete_test":"Retrain D5C5 from scratch on exactly the same 100 training volumes used for LANTERN, with the same cropping, the same undersampling masks, and the same 50 test volumes, at 1D random 4x and 7x acceleration. If D5C5's PSNR/SSIM/HFEN on that test set is within roughly 0.5-1 dB PSNR of LANTERN's Table 4 values, the claimed small-data advantage is not supported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's core claim is that LANTERN reconstructs dynamic cardiac MRI better than deep-learning methods when training data are limited. The only deep-learning comparator, D5C5, is trained on 2900 volumes while LANTERN uses 100 (Implementation section). The text concedes: 'if we use 100 data to train, the reconstruction results are not good enough, and the experimental results also verify the conjecture,' but no same-size D5C5 experiment is reported or shown. Without that control, the comparison conflates architecture advantage with training-set-size advantage: D5C5 is evaluated outside its intended data regime, and LANTERN may simply benefit from being better tuned for small data. Additionally, the 'Impact of dataset size' section selects N=100 because it gives the best PSNR on the 50 test volumes (Table 2), so the reported LANTERN number is selected on the test set. The central claim therefore rests on an assumption about D5C5's small-data performance that is asserted but not demonstrated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LANTERN, an unrolled ADMM network for dynamic MRI reconstruction that learns analysis sparsifying transforms via a small set of training volumes. The network is initialized with DCT (spatial) and TV (temporal) filters, unrolls a fixed number of ADMM stages, and replaces the derivative of the prior with a learned piecewise-linear function. On 1D random and 2D radial undersampling masks, the authors report higher PSNR, SSIM, and HFEN than k-t SLR and D5C5 across acceleration factors from 2x to 15x, training with 100 cardiac volumes. The paper claims that this demonstrates superior reconstruction accuracy when the amount of training data is limited.","tokens_in":14458,"tokens_out":3562,"duration_ms":37130,"significance":"If the small-data claim is established, the result would be practically significant for clinical MRI, where large fully-sampled training sets are often unavailable. The unrolled architecture is a credible model-based deep learning approach, and the paper provides useful details: quantitative metrics with standard deviations, a comparison of initialization schemes, and training/validation loss curves. The attempt to combine a learnable analysis prior with DCT/TV initialization is a reasonable design choice. However, the central claim is currently supported only by an indirect comparison: the deep-learning baseline D5C5 is trained on 2900 volumes, not on the same 100 volumes used for LANTERN, and the reported LANTERN numbers are partially selected on the test set. The significance is therefore conditional on a matched evaluation.","major_comments":[{"comment":"The paper's core claim that LANTERN is superior to deep-learning baselines under small training data is not directly tested. D5C5 is trained on 2900 volumes (300 testing) while LANTERN is trained on 100 volumes and tested on 50. The text asserts that with 100 training data 'the reconstruction results are not good enough' and that experiments `verify the conjecture', but no experiment with D5C5 trained on the same 100 volumes is reported or shown. As written, the comparison conflates architecture advantage with training-set-size advantage; if D5C5 also performs well at 100 training volumes, the claimed small-data advantage is not established. Reporting this matched control is essential.","section":"Implementation (paragraph on D5C5); Tables 4 and 5"},{"comment":"The training-set size and the initialization scheme are selected using the 50 test volumes. Table 2 reports reconstruction metrics on the test set for training sizes 50, 60, 80, 100, 120, and the text says 'we selected 100 images' because the test PSNR is best at 100. Similarly, Table 3 compares Random Gaussian, DCT, and DCT+TV initializations on the same 50 test volumes and chooses DCT+TV. This makes the headline numbers in Tables 4 and 5 partially test-selected, so they are not a clean held-out evaluation. A validation split (or nested selection) should be used for these choices, and the final test numbers should be reported after the selection is fixed.","section":"Impact of dataset size (Table 2, Fig. 3); Effect of initializations (Table 3)"},{"comment":"The data description is internally inconsistent and matters for the small-data claim. The Dataset section says 101 fully sampled cardiac datasets were collected and that '150 cardiac data were generated with 100 data for network training, and 50 for network testing', while the Implementation section says the data were cut to 126x126x16 and 3200 data were obtained, with 2900 for D5C5 training and 300 for testing. It is unclear how 101 datasets become 150, how many distinct subjects or slices are in the 100/50 split, and whether the 2900 D5C5 volumes overlap with the 100 LANTERN volumes. Please clarify this pipeline and, ideally, evaluate LANTERN on the same 300-volume test set used for D5C5, or evaluate D5C5 on the same 50-volume test set.","section":"Dataset paragraph and Implementation paragraph"},{"comment":"The expression for d x^(n) / d rho^(n) appears to use a scalar-style derivative of a matrix inverse, writing terms like (P^H P + rho I)^{-2} with the inverse raised to a power. As written this is not a valid matrix derivative. Since the end-to-end training of the network is a central contribution, the gradient derivation for the rho parameters should be either corrected, stated with proper operator notation (e.g., - (P^H P + rho I)^{-1} (d/dt)[P^H P + rho I] (P^H P + rho I)^{-1}), or justified by reference to a standard matrix calculus result.","section":"Back-propagation, ReCon layer gradient (paragraph after Eq. (8))"}],"minor_comments":[{"comment":"The text says the loss is the standard mean square error, but the equation for E(Theta) defines a normalized RMSE loss (square root of a normalized sum of squared errors). Please align the text with the actual loss.","section":"Loss definition"},{"comment":"There is a typo, 'undersamling', in the abstract.","section":"Abstract"},{"comment":"The notation in Eqs. (7)-(8) is inconsistent: the Addition layer uses v^(n,k-1) while the multiplier update uses v^(n), and the warm-start cases for n=1 are written separately but not clearly tied to the general formulas. Please unify the superscripts and state explicitly which v is passed to the Multi layer.","section":"Eq. (8) and layer notation"},{"comment":"The sentence 'the size of the applied filter is 3*3*9, where the first eight (3*3*8) are DCT and the last one (3*3*1) is TV' is unclear: please specify how the temporal dimension is arranged and how the 3D convolution and 2D TV filter are combined.","section":"Filter description"},{"comment":"The convergence analysis only shows that training and validation losses decrease. This is evidence against overfitting, but it does not by itself demonstrate that the unrolled network converges to a solution of the original optimization problem. Consider adding a plot of reconstruction error on a fixed validation set per epoch, or a comparison with the optimization-based solution at initialization.","section":"Convergence analysis (Fig. 10)"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim is the small-data advantage over data-hungry deep methods, and that claim is currently not supported by a matched comparison. This is fixable within the scope of the manuscript: train D5C5 (or another deep baseline) on the same 100 volumes and report the same metrics, and use a validation set for model selection. The theoretical derivation of the unrolled network is plausible but should be checked carefully, particularly the rho gradient. If the matched comparison still shows clear gains for LANTERN with 100 training volumes, the paper would be a solid contribution to the dynamic MRI reconstruction literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the method itself is a reasonable and clearly explained unrolled ADMM network for dynamic MRI, with the small twist of initializing the 3D filters as DCT in space and TV in time and then training end-to-end. Second, the headline claim—that it beats a deep cascade (D5C5) when training data are limited—is not actually tested. D5C5 is trained on 2900 volumes while LANTERN uses 100, and the paper explicitly concedes that D5C5 with 100 volumes \"reconstruction results are not good enough\" but it never shows that experiment. That is a load-bearing gap, not a minor omission.\n\nWhat is genuinely good: the derivation from the ADMM optimization to the network layers is explicit, with forward and backward formulas spelled out. The experimental tables include standard deviations and cover a useful range of acceleration factors (2x-11x 1D random, 2x-15x 2D radial) and two undersampling patterns. The comparison to k-t SLR is informative, and the reported reconstruction time under 3 seconds is practically relevant. The paper is honest enough to state the small-data conjecture in the Implementation section, which makes the missing control more frustrating rather than deceptive.\n\nThe soft spots are real but fixable. Beyond the missing same-data-size D5C5 comparison, the \"impact of dataset size\" section selects 100 training volumes because it gives the best PSNR on the 50 test volumes, and the initialization is also chosen by test-set performance. So the reported numbers are partially selected on the test set; a proper validation split is needed. No code or data is released, which matters for a paper whose main claim is empirical. These are not fatal to the architecture, but they do mean the central practical advantage is unproven.\n\nWho is this for? Researchers working on dynamic MRI reconstruction with limited datasets, especially those interested in unrolled optimization networks. They will get a useful baseline and a clear derivation, but they should not treat the small-data superiority claim as established. The paper deserves a serious referee: the problem is important, the method is coherent, and the missing experiments are straightforward to request. My recommendation: send it to peer review, but with the explicit expectation that the authors add a same-size D5C5 comparison, a clean validation split, and ideally code/data release. If they cannot show the small-data advantage under a matched training budget, the paper should still be publishable as an incremental unrolled-network contribution, but not with the current marketing.","headline":"A plausibly useful unrolled ADMM network for dynamic MRI, but the central small-data claim is asserted rather than demonstrated because D5C5 is never trained on the same 100 volumes.","tokens_in":15083,"tokens_out":1123,"would_cite":false,"duration_ms":13346,"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":"LANTERN, a deep network that unrolls ADMM with DCT and TV priors, reconstructs dynamic cardiac MRI better than k-t SLR and D5C5 using only 100 training volumes.","keywords":["dynamic MRI reconstruction","cardiac cine MRI","compressed sensing","ADMM unrolling","deep learning","small training dataset","k-space undersampling","analysis transform learning"],"falsifier":"Train D5C5 (or an equivalent data-hungry deep reconstruction network) on the same 100 training volumes used for LANTERN, with identical test data and sampling masks, and compare PSNR, SSIM, and HFEN; if the retrained D5C5 matches or beats LANTERN's reported 41.44 dB PSNR at 1D random 4x acceleration, the central small-data advantage claim fails.","tokens_in":14023,"feed_emoji":"🧲","tokens_out":7216,"duration_ms":65715,"temperature":0.7,"pith_summary":"This paper tries to establish that a model-based deep network, LANTERN, can reconstruct dynamic cardiac MRI from heavily undersampled k-space data using only 100 training volumes, while outperforming both the compressed-sensing baseline k-t SLR and the large-data deep baseline D5C5. The significance is that dynamic MRI acceleration currently faces a trade-off: classical compressed sensing needs hand-tuned parameters, while deep reconstruction methods typically require thousands of training samples. LANTERN unrolls an ADMM solver of a sparse convolutional coding model into network layers and learns all parameters end-to-end. On 1D random (2x-11x) and 2D radial (2x-15x) undersampling, it reports higher PSNR, SSIM, and HFEN than both baselines, with reconstruction in under three seconds. If correct, this would make high-quality accelerated dynamic MRI practical in settings where large training datasets are unavailable.","feed_headline":"100 training volumes beat big-data MRI reconstruction networks","feed_subtitle":"An ADMM-unrolled net with DCT+TV priors tops k-t SLR and D5C5 at up to 15x acceleration.","key_machinery":"The load-bearing object is the unrolled ADMM iteration, expressed as six layer types: Recon, Addition, Conv1, Nonlinear, Conv2, and Multi. Conv1 and Conv2 learn convolutional analysis operators that sparsify the image, the Nonlinear layer approximates the derivative of the regularization function by a piecewise linear function, and the Recon layer performs the closed-form k-space data-consistency update. The DCT and TV initialization provides the prior structure that lets the network work from a small training set.","core_discovery":"LANTERN treats dynamic MR reconstruction as a sparse convolutional coding problem with a data-fidelity term and learned analysis transforms, and solves it by unrolling alternating direction method of multipliers (ADMM) iterations into a fixed-depth network. The spatial transform is initialized with the discrete cosine transform (DCT) and the temporal transform with total variation (TV), and every parameter—filters, regularization weights, step sizes, and the piecewise-linear nonlinearity—is learned end-to-end from undersampled-to-fully-sampled pairs. On cardiac data, the trained network reaches higher average PSNR, SSIM, and HFEN than k-t SLR and D5C5 at every tested acceleration factor, including 11x random and 15x radial undersampling, while training on 100 volumes versus 2900 for D5C5.","pith_inferences":["A decisive test the paper does not run is to train D5C5 on the same 100 volumes; if that retrained network also reconstructs well, the claimed small-data advantage would be an artifact of comparison size rather than of the LANTERN architecture.","The DCT+TV initialization likely acts as an inductive bias that reduces data demand; ablating the initializer while keeping the rest of the architecture fixed would isolate how much of the small-data performance comes from initialization versus learned layers.","The same recipe—unrolling an ADMM solver and initializing analytic transforms—could transfer to other inverse problems with known sparsifying bases, such as CT or PET, but the paper does not test this."],"forward_implications":["If the central claim holds, dynamic cardiac MRI can be reconstructed accurately from 2x-11x random or 2x-15x radial undersampling with a network trained on 100 volumes, far fewer than the thousands used by the D5C5 baseline.","The unrolled ADMM structure preserves a compressed-sensing interpretation: the learned filters act as data-adaptive sparsifying transforms, and the data-fidelity update remains explicit.","Reconstruction time under three seconds, versus about 200 seconds for k-t SLR, would make the approach clinically practical despite a training phase measured in hours.","Because all regularization and step-size parameters are learned, the manual tuning burden of classical CS-MRI reconstruction is removed."],"supporting_citations":[{"why":"This is the D5C5 deep-cascade baseline that LANTERN must beat on small data; the paper trains it on 2900 volumes and compares against it.","marker":"[4]"},{"why":"This is the k-t SLR compressed-sensing baseline exploiting sparsity and low-rank structure that LANTERN reports higher PSNR, SSIM, and HFEN than.","marker":"[13]"},{"why":"This is the Deep ADMM-Net that supplies the unrolled-optimization architecture LANTERN adapts into its reconstruction, prior, and multiplier layers.","marker":"[23]"}],"fun_headline_variants":["100 volumes beat 2900 for cardiac MRI reconstruction","LANTERN: ADMM-unrolled net tops k-t SLR with 100 training sets","MRI from 100 samples: LANTERN beats D5C5 trained on 2900","Small dataset MRI: 100 volumes outperform big-data networks","Learn analysis transforms: better reconstruction with less data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes D5C5 is a valid control at the same training size, but the paper trains D5C5 on 2900 volumes and LANTERN on 100, asserting without shown evidence that D5C5 performs poorly with 100 volumes.","fun_headline_variants_meta":{"raw":{"variants":["100 volumes beat 2900 for cardiac MRI reconstruction","LANTERN: ADMM-unrolled net tops k-t SLR with 100 training sets","MRI from 100 samples: LANTERN beats D5C5 trained on 2900","Small dataset MRI: 100 volumes outperform big-data networks","Learn analysis transforms: better reconstruction with less data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000417,"raw_usage":{"total_tokens":2126,"prompt_tokens":897,"completion_tokens":1229,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":1137}},"tokens_in":513,"tokens_out":1229,"duration_ms":9929,"temperature":1.0,"reasoning_tokens":1137,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:20:43.643249+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train D5C5 (or an equivalent data-hungry deep reconstruction network) on the same 100 training volumes used for LANTERN, with identical test data and sampling masks, and compare PSNR, SSIM, and HFEN; if the retrained D5C5 matches or beats LANTERN's reported 41.44 dB PSNR at 1D random 4x acceleration, the central small-data advantage claim fails.","supporting_citations":[{"cited_title":"A deep cascade of convolutional neural networks for dynamic MR image reconstruction,","cited_arxiv_id":null,"evidence_quote":"This is the D5C5 deep-cascade baseline that LANTERN must beat on small data; the paper trains it on 2900 volumes and compares against it."},{"cited_title":"Accelerated dynamic MRI exploiting sparsity and low-rank structure: K-t SLR,","cited_arxiv_id":null,"evidence_quote":"This is the k-t SLR compressed-sensing baseline exploiting sparsity and low-rank structure that LANTERN reports higher PSNR, SSIM, and HFEN than."},{"cited_title":"Deep ADMM-Net for compressive sensing MRI,","cited_arxiv_id":null,"evidence_quote":"This is the Deep ADMM-Net that supplies the unrolled-optimization architecture LANTERN adapts into its reconstruction, prior, and multiplier layers."}],"review_version":1}