{"id":"113b3487-80a5-4fce-a808-991ff10ef43d","arxiv_id":"1908.05615","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A multi-sequence MRI recovery network trained on random undersampling masks, then fine-tuned on a chosen strategy, outperforms recovering each sequence independently and offers a cheap way to pick sampling factors.","lead":"This paper studies how to recover multiple MRI scan sequences at once from undersampled data, and how to divide a fixed scan time budget among the sequences. The proposed neural network and search strategy could make multi-sequence MRI faster while keeping images usable.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (5) search is validated only on a post hoc subrange of the test set; test-set-selected Table 1 numbers may inflate both search quality and MIMO-tuned gains.","rationale":"The reader's weakest assumption flags the surrogate-ranking risk in Eq. (5): the BRM's loss landscape may not rank strategies the same way dedicated models would, and the validation is restricted to a post hoc selected range. I agree that this is the most load-bearing assumption, but I would sharpen it: the experiments also use the same test set for both selecting the strategies and reporting the final numbers, which creates a selection-on-test-set bias. The paper does not report error bars, code, or data, and the private test set is small (roughly 288 three-sequence images from two patients after the split), so the statistical basis for the headline gap is thin. Even so, the MIMO-vs-SISO comparison in Table 1 is consistently in favor of MIMO even without fine-tuning, so the recovery claim is not obviously false. The conditional verdict remains appropriate: the paper's central search claim needs a validation-set re-run and a fair fine-tuned SISO baseline before the reported gains can be trusted. No fraud or bad faith is implied; the issue is an experimental-protocol gap that a clean split would resolve.","tokens_in":7720,"tokens_out":7273,"duration_ms":69956,"concrete_test":"Use only the validation fold for all selection: after training the BRM on the training fold, evaluate Eq. (5) on validation data over the full lambda-simplex, train dedicated models for a pre-specified set of 50 candidate strategies (including the BRM top-10), and compute Spearman rank correlation and top-3 overlap between BRM losses and dedicated-model losses. Then fine-tune MIMO and, for fairness, SISO models on the validation-selected top-3 strategies, and evaluate once on the untouched test fold. If the selected lambdas differ from Table 1, or the test PSNR advantage of MIMO-tuned over fine-tuned SISO drops below about 0.5 dB, then the current Table 1 numbers are a test-set artifact and Eq. (5) is not shown to be a reliable surrogate.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is Eq. (5): the blind recovery model's loss is used to rank sampling strategies, and the top-ranked strategies are then fine-tuned and reported in Table 1. For this to support the paper's efficient-search claim, the BRM loss must rank strategies essentially the same way a dedicated model for each strategy would. The paper's only evidence for this is Fig. 2: after plotting BRM and ZF performance on the test set, the authors select a subrange of sampling factors where ZF correlates poorly, train 30 dedicated models in that subrange, and report a Pearson correlation of 0.85 (MIMO) for that subrange. This is post hoc: the subrange was chosen after seeing the test-set scatter, and the same test set is used to select the top-3 strategies and to produce the final PSNR/SSIM numbers. Pearson correlation on a selected subrange does not establish top-K ranking fidelity, and reporting numbers from strategies selected on the test set is a selection-on-test-set bias. The MIMO-tuned gains in Table 1 may therefore be inflated and the chosen lambdas may be suboptimal for dedicated or even fine-tuned SISO models. If the search is unreliable, the central claim of efficient strategy selection is unsupported; the MIMO-vs-SISO recovery advantage for fixed masks is less affected, but the headline pipeline includes both search and recovery.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses multi-sequence MR image recovery from undersampled k-space data under an overall acquisition-time budget. It formulates a constrained optimization over sampling masks and a recovery network (Eq. (2)), argues that joint optimization is combinatorial, and proposes a blind recovery model (BRM) trained with randomly sampled masks. The trained BRM is then used as a surrogate to search for good sampling strategies via Eq. (5), and the selected strategy is used for fine-tuning via Eq. (6). The authors compare a single multi-input multi-output (MIMO) BRM with per-sequence single-input single-output (SISO) models and with dedicated per-strategy models, on a private real T1/T2/FLAIR dataset and on simulated BraTS data. Table 1 reports the top three strategies found by the search for several acquisition-time settings and shows consistent MIMO-over-SISO PSNR/SSIM improvements.","tokens_in":8025,"tokens_out":8872,"duration_ms":82182,"significance":"If the results hold, the paper makes a useful contribution: it is one of the first to cast multi-sequence undersampling as a joint optimization of masks and recovery network, and the idea of using a randomly trained BRM as a surrogate to avoid training many dedicated models is practically appealing. The paper's strengths include the explicit constrained formulation, the effort to validate the surrogate with dedicated models, and the consistent MIMO-over-SISO advantage across the twelve rows of Table 1. However, the evaluation currently has serious statistical and selection issues: the search surrogate is validated only on a post hoc selected subrange, the top strategies are chosen and evaluated on the same test set, and the real-data test set is very small. These issues affect the central claim that the proposed search reliably identifies good sampling strategies, so the paper needs substantial revision before the conclusions can be accepted.","major_comments":[{"comment":"The load-bearing claim is that the loss of the blind recovery model in Eq. (5) ranks sampling strategies the way dedicated models would. The only evidence is the Pearson correlation coefficient of 0.85 between \"Dedicated\" and \"MIMO\" in Fig. 2 (versus -0.33 for ZF), but that coefficient is computed on a subrange of sampling factors the authors selected after plotting the test-set scatter, specifically where ZF performance \"does not correlate well.\" A correlation on a post hoc subrange does not establish top-K ranking fidelity, which is what Eq. (5) requires if the top three strategies in Table 1 are to be trusted. Please pre-specify the candidate range, report Spearman correlation or top-K agreement over the full candidate set, and evaluate the ranking on a validation set rather than the test set.","section":"§3 (Main results), Eq. (5), Fig. 2"},{"comment":"The top-three sampling strategies in Table 1 are selected by running the trained models on the test set (the text says \"We run the trained models on the test set, and plot the reconstruction performances... top-three performing sampling strategies\"), and the same test-set numbers are then reported as the results. This is selection on the test set: the reported PSNR/SSIM values are maxima over hundreds of candidates and are therefore optimistically biased. Please perform the selection on the validation split and use the test set only for final evaluation; if that is not possible, explicitly quantify the selection bias by reporting, for example, the mean and maximum performance across the candidate grid.","section":"§3 (Main results), Table 1"},{"comment":"No error bars or significance tests are provided. The real-data split is 17:1:2 on 20 patients, so the real-data evaluation rests on about two patients even after treating each channel as a separate image, because patients are the independent units. The MIMO-vs-SISO differences in Table 1, such as 39.24 vs 38.57 dB in the first row, are small and could be within inter-patient variability. Please report per-patient or per-slice standard deviations and bootstrapped confidence intervals, and state clearly how many patients are in the test set.","section":"§3 (Datasets; Main results)"},{"comment":"Because the same BRM is used as the search surrogate in Eq. (5) and as the initialization for fine-tuning in Eq. (6), the final \"MIMO (tuned)\" numbers are not independent evidence that the selected masks are near-optimal for a model trained specifically for those masks. The comparison with 30 dedicated models is the appropriate control, but it is computed only in the post hoc subrange discussed above. Please report dedicated-model PSNR/SSIM for the exact λ values shown in Table 1, or otherwise justify that fine-tuning from the BRM attains dedicated-model performance.","section":"§2.3, Eq. (6)"}],"minor_comments":[{"comment":"The notation fθ(IMs) is inconsistent with the multi-sequence architecture shown in Fig. 1, where the network takes all three undersampled images as input. Please write the input as a tuple, e.g., fθ(IM1,...,IMS), or explicitly define that MIMO inputs are implicit.","section":"Eqs. (2) and (4)"},{"comment":"The sampling distribution p(C) is not defined. The text says λs is sampled randomly from [1,k], but the candidate set C used in Eq. (5) and the measure over masks should be specified precisely, since the search cost depends on |C|.","section":"§2.1, Eq. (4)"},{"comment":"The choice of random versus low-pass sampling is made per dataset after seeing which performs better. Please state whether this choice is fixed before the search procedure and whether the mask family is held fixed across all compared strategies; otherwise the search space is not fully specified.","section":"§3 (Acquisition time and undersampling settings)"},{"comment":"Because the L1 objective in Eq. (2) sums across sequences with different dynamic ranges, the search may be dominated by high-intensity sequences such as T2. Please report per-sequence PSNR/SSIM for the selected strategies and discuss whether equal weighting of the sequence losses is appropriate.","section":"§3 (Evaluation metrics)"},{"comment":"The supplemental figures all carry the caption \"Fig. 2\"; please renumber them and refer to them correctly from the main text.","section":"Supplemental Material"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the MIMO-vs-SISO recovery advantage appears consistent, so I do not recommend rejection. However, the search validation and the test-set selection issues are load-bearing, and the real-data test set is very small and private. Encouraging the authors to release code and, if possible, a larger or public test set would substantially increase the credibility of the empirical claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Ho — quick take on 1908.05615. The paper is worth a look, but with a clear-eyed view.\n\nWhat's actually new: they formulate multi-sequence MRI acceleration as a joint optimization over sampling masks and a recovery network, and they make it tractable by training one blind network on random masks, using that network's loss to rank candidate sampling strategies, and then fine-tuning on the top-ranked one. The random-mask training for a multi-sequence MIMO network is a reasonable extension of existing single-sequence blind reconstruction work. The consistent MIMO-over-SISO gains (roughly 0.5–0.9 dB PSNR across settings, before tuning) are credible because they show up across four different acquisition-time setups and on both real and simulated data. That part of the paper I believe.\n\nThe search-validity part is softer. The evidence that BRM loss ranks strategies like dedicated models is Fig. 2, a Pearson correlation of 0.85 — but that correlation is computed on a subrange of sampling factors selected after looking at the test-set scatter. The top strategies in Table 1 are also chosen on the test set before fine-tuning, so the reported numbers are likely optimistic. There are no error bars, the real-data test set is two patients, and they picked random vs low-pass sampling per dataset after the fact. None of these kill the recovery claim, but they weaken the 'efficient search' claim: we don't actually know how close the found strategy is to the true optimum for a dedicated model.\n\nI'd send it to peer review — the problem is well posed and the idea is worth discussing — but I'd push for error bars, a held-out validation of the search (e.g., train dedicated models on strategies not used to pick the top-K), and code/data release. If the authors can tighten that, the paper becomes a useful reference for multi-sequence MRI.","headline":"Solid multi-sequence MRI recovery paper with a novel blind-search idea, but the search validation is post hoc and the test set is too small to fully trust the efficiency numbers.","tokens_in":8511,"tokens_out":3042,"would_cite":false,"duration_ms":31350,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a single MIMO blind recovery model, trained on randomly sampled masks and then fine-tuned on the selected masks, recovers multiple MR sequences better than independently recovering each sequence, and that the model's…","keywords":["multi-sequence MRI","blind recovery model","undersampled k-space","sampling strategy search","deep learning","image reconstruction","MIMO network","time budget"],"falsifier":"Train dedicated models for the strategy the BRM search ranks best and for the one it ranks worst under a fixed time budget; if the dedicated model performs worse on the BRM-best strategy than on the BRM-worst strategy, the surrogate ranking fails. A cheaper check: compare BRM-predicted losses against the measured PSNR of dedicated models across several strategies, which must be monotonically related for the paper's claim to hold.","tokens_in":7523,"feed_emoji":"🧲","tokens_out":6635,"duration_ms":53638,"temperature":0.7,"pith_summary":"MRI scans routinely acquire several sequences (T1, T2, FLAIR) for the same patient, and each sequence takes time; undersampling k-space speeds things up but degrades images. The paper treats the joint choice of how many phase-encoding lines to keep for each sequence and the recovery network as a constrained optimization under a fixed scan-time budget. It shows that naively searching over all sampling strategies and retraining a network for each is combinatorial, and proposes a \"blind recovery model\": one multi-input, multi-output network trained on randomly undersampled versions of all sequences at once. The central claims are that this single network recovers all sequences better than per-sequence networks, and that its loss can rank candidate sampling strategies cheaply, so a near-optimal strategy can be found and the model fine-tuned on it.","feed_headline":"Blind CNN finds best undersampling plan for multi-sequence MRI","feed_subtitle":"Trained on random masks, a multi-input network beats per-sequence recovery and finds how to spend a fixed scan time.","key_machinery":"The load-bearing mechanism is the blind recovery model (BRM): a MIMO CNN built from Residual Dense Blocks that takes the undersampled images of all sequences, concatenated along the channel dimension, and outputs the recovered images for all sequences at once. It is trained by data augmentation over masks, drawing each sequence's undersampling factor λ uniformly from [1, k], so the model learns to handle a continuum of sampling profiles rather than one fixed mask. This trained BRM is then reused as a loss-based oracle: exhaustive search over candidate mask combinations (Eq. 5) requires only forward passes, not retraining, after which the chosen masks are used to fine-tune the model (Eq. 6). The dense residual connections let the network share anatomical information across sequences.","core_discovery":"On the paper's own terms, the discovery is that a MIMO blind recovery model—trained once with masks drawn randomly across a continuous range of undersampling factors on every sequence—serves both as a multi-sequence reconstructor and as a surrogate quality estimator for sampling-strategy search. Because the model has seen many masks, its loss on a candidate strategy predicts the performance of a dedicated model trained specifically for that strategy (the reported Pearson correlation is 0.85, versus -0.33 for the zero-filled input, in the tested range). Using this loss, the paper selects a near-optimal set of per-sequence undersampling factors under a fixed time budget (for example, 75% time reduction), then fine-tunes the model on those masks. Across real and simulated three-sequence data, the resulting pipeline beats recovering each sequence independently, and the chosen strategy consistently samples T2 most heavily.","pith_inferences":["Since the surrogate ranking only needs forward passes, the same scheme could optimize other acquisition parameters (e.g., number of excitations or slice spacing) provided the mask distribution covers them.","The method's reliance on co-registered sequences suggests a testable boundary: if sequences are not aligned, the benefit of MIMO over SISO may shrink, because the network's cross-sequence sharing depends on shared anatomy.","A natural next test is to run the BRM loss on low-resolution scout images of a specific patient to personalize the sampling strategy on the fly, rather than using a fixed policy found on a population."],"forward_implications":["A single trained model replaces the need to train a separate network for each candidate sampling strategy, cutting the search cost from training $F^S$ models to $F^S$ forward passes.","The MIMO model consistently outperforms per-sequence SISO models on both real and simulated data in the reported settings, so multi-sequence recovery can be done with one network instead of several.","The search procedure yields an interpretable allocation rule: within a time budget, the model tends to allocate more samples to the sequence with the best contrast (T2), and reallocates toward T1 when T1 acquisition is fast (time ratios 2:3:6).","The framework is agnostic to the underlying sampling pattern (low-pass or random) and extends beyond three sequences by treating additional sequences like extra image channels."],"supporting_citations":[{"why":"Shows that complementary T1 information can help reconstruct a highly undersampled T2 image, motivating the cross-sequence sharing used by the MIMO model.","marker":"[12]"},{"why":"Defines the compressed-sensing MRI setup with phase-encoding-line masks, which the paper adopts for undersampling.","marker":"[1]"},{"why":"Supplies the Residual Dense Block architecture that the blind recovery network is built from.","marker":"[13]"},{"why":"Provides the BraTS benchmark dataset used to create simulated multi-sequence MRI data for validation.","marker":"[15]"},{"why":"Provides the expert-labeled glioma MRI collection that the simulated dataset is derived from.","marker":"[16]"}],"fun_headline_variants":["Blind network sets per-sequence undersampling for MRI time budget","One CNN recovers multi-sequence MRI and picks sampling strategy","Mask-trained blind model outperforms per-sequence MRI recovery","Multi-sequence MRI recovery with automatic undersampling plan","Blind CNN predicts best MRI sampling under time constraint"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The search step assumes that a model trained on random masks ranks candidate sampling strategies in the same order that models dedicated to each strategy would; the paper validates this only on a post hoc range of sampling factors, so the top-ranked strategy could be suboptimal for the dedicated model.","fun_headline_variants_meta":{"raw":{"variants":["Blind network sets per-sequence undersampling for MRI time budget","One CNN recovers multi-sequence MRI and picks sampling strategy","Mask-trained blind model outperforms per-sequence MRI recovery","Multi-sequence MRI recovery with automatic undersampling plan","Blind CNN predicts best MRI sampling under time constraint"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00019,"raw_usage":{"total_tokens":1305,"prompt_tokens":876,"completion_tokens":429,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":492,"completion_tokens_details":{"reasoning_tokens":351}},"tokens_in":492,"tokens_out":429,"duration_ms":4790,"temperature":1.0,"reasoning_tokens":351,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:07:38.986375+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train dedicated models for the strategy the BRM search ranks best and for the one it ranks worst under a fixed time budget; if the dedicated model performs worse on the BRM-best strategy than on the BRM-worst strategy, the surrogate ranking fails. A cheaper check: compare BRM-predicted losses against the measured PSNR of dedicated models across several strategies, which must be monotonically related for the paper's claim to hold.","supporting_citations":[{"cited_title":"In: MICCAI","cited_arxiv_id":null,"evidence_quote":"Shows that complementary T1 information can help reconstruct a highly undersampled T2 image, motivating the cross-sequence sharing used by the MIMO model."},{"cited_title":"Magnetic Resonance in Med","cited_arxiv_id":null,"evidence_quote":"Defines the compressed-sensing MRI setup with phase-encoding-line masks, which the paper adopts for undersampling."},{"cited_title":"IEEE Trans","cited_arxiv_id":null,"evidence_quote":"Provides the BraTS benchmark dataset used to create simulated multi-sequence MRI data for validation."},{"cited_title":"Scientific data 4 (2017) 10 Cheng Peng 1, Wei-An Lin1, Rama Chellappa 1, S","cited_arxiv_id":null,"evidence_quote":"Provides the expert-labeled glioma MRI collection that the simulated dataset is derived from."}],"review_version":1}