{"id":"25ea9a93-a7fe-4695-8b94-92d8fdeeba7d","arxiv_id":"2607.19779","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"HieraSample, a Mamba-based frequency-hierarchical active sampler, matches fully-sampled ACL diagnosis AUC at 4-10x acceleration on fastMRI+ knee MRI.","lead":"HieraSample is an MRI sampling method that decides, step by step, which k-space points to measure: it always keeps the low-frequency center of k-space and lets a Mamba-based policy pick high-frequency points to maximize diagnostic accuracy. On fastMRI+ knee data it reports matching fully-sampled ACL diagnosis at 4-10x acceleration and large AUC gains over an existing Cartesian baseline.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reward and evaluation use the same frozen classifier; reported oracle parity and +20.4 AUC may reflect policy exploiting classifier-specific loss surface rather than diagnostic information.","rationale":"The reader's weakest_assumption correctly identifies the shared classifier between reward and evaluation as the load-bearing premise. This is indeed the single most critical soft spot: without an independent classifier or human read, the headline parity and large AUC gains could be an artifact of policy overfitting to the frozen classifier's loss surface. The proposed test—evaluating the learned masks with a second, architecturally distinct classifier—would directly settle whether the gains reflect diagnostic information or classifier-specific exploitation. Since the reader already flagged this and conditioned acceptance on such validation, the verdict should remain CONDITIONAL rather than being moved to REJECT or ACCEPT. The other concerns (single-seed noise, non-monotonic baseline) are real but secondary; the circularity is the most decisive.","tokens_in":7365,"tokens_out":3575,"duration_ms":35509,"concrete_test":"Train a second, architecturally distinct classifier (e.g., a standard ResNet-50 without Mamba or a ViT) on fully-sampled fastMRI+ knee data using the same train/val/test split. Feed the zero-filled reconstructions generated by the HieraSample masks at each acceleration to this independent classifier and compute ROC AUC for ACL diagnosis and severity. If the gap to the fully-sampled oracle widens beyond the ±0.003/±0.015 bootstrap intervals, or the +20.4 point gain over ASSDM shrinks by more than half, the reported results are partly an artifact of the shared classifier reward/evaluation loop.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on a closed loop between training signal and evaluation. In §3.5, the reward r_t is the per-step reduction in class-weighted cross-entropy of the frozen disease/severity classifiers (Eq. 3). The headline results in §4.2 (Table 1) report ROC AUC of those same frozen classifiers on the zero-filled reconstructions produced by the policy. Because the policy is trained with REINFORCE to maximize this exact CE reduction, it is rewarded for making the classifier more confident, not necessarily for adding clinically relevant information. If the classifier's confidence is miscalibrated or its loss surface has idiosyncratic 'easy' directions, the policy can exploit them; the reported 'matches fully-sampled oracle' at 10x/8x/6x and the +20.4 AUC gain on ACL severity would then not transfer to an independent classifier, a different reconstruction algorithm, or a radiologist. The paper does not validate with a second classifier or human read. The conclusion acknowledges single-seed and single-anatomy limitations, but does not address this circularity, and the condition that the classifier be jointly fine-tuned on the policy's mask distribution (mentioned as future work) is a symptom of the same issue: the frozen classifier's CE may be an unreliable reward oracle. The 'matches' cells are also within the reported single-seed bootstrap interval, but the circularity is the more fundamental threat.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents HieraSample, a task-driven active k-space sampling method for accelerated MRI. It combines a cosine-annealed acceleration curriculum over 80 steps, a preserved fully-sampled low-frequency disk, and a Mamba-based policy that selects individual high-frequency k-space coordinates. The reward is the per-step reduction in class-weighted cross-entropy of frozen disease and severity classifiers. On fastMRI+ knee, the authors report that HieraSample matches the fully-sampled oracle on ACL diagnosis at 4x-10x acceleration and improves on a retrained ASSDM-Cartesian baseline by up to 20.4 AUC points on ACL severity. The evaluation is single-seed, uses the same frozen classifiers for reward and evaluation, and does not release code or learned masks.","tokens_in":7614,"tokens_out":4802,"duration_ms":46851,"significance":"The paper addresses a practically important question: where to spend a limited k-space budget in a task-driven way. The frequency-hierarchical decomposition and point-based action space are sensible design choices, and the use of Mamba for 80-step rollouts is an interesting architectural contribution. The comparison with a Cartesian action-space ablation under the same classifier and reward helps isolate the contribution of the action space. However, because the reward and the headline metric are computed with the same frozen classifiers, and because the reported oracle parity is drawn from a single seed with bootstrap intervals that cover the gaps, the results should be treated as promising but not yet conclusive. The limitations in §5 are candid about scope but do not resolve the central evaluation concern.","major_comments":[{"comment":"The reward r_t is the per-step reduction in class-weighted cross-entropy of the frozen classifiers G_d and G_s, and the headline AUC in Table 1 is computed with those same frozen classifiers. REINFORCE optimizes the policy to maximize this CE reduction, so the evaluation is not independent of the training signal. A policy can reduce CE by making the classifier more confident on its own loss surface without improving the information content of the mask; the reported 'matches fully-sampled oracle' and +20.4 AUC gain would then not transfer to another classifier, another reconstruction, or a radiologist. The conclusion's statement that joint fine-tuning of the classifier on the policy's mask distribution is future work (§5) confirms that the current setup is a closed loop. I would want either an evaluation with a second, independently trained classifier (preferably a different architecture)","section":"§3.5, Eq. (3), §4.2"},{"comment":"The abstract and §4.2 claim HieraSample 'matches the fully-sampled oracle on ACL diagnosis from 4x to 10x acceleration.' With a single seed and paired-bootstrap intervals of ±0.003 AUC (and ±0.015 for ACL severity), several supporting cells are statistically indistinguishable from the oracle rather than independent confirmations: e.g., ACL diagnosis at 4x is 0.938 vs. oracle 0.939, and ACL severity at 6x is 0.838 vs. 0.837. The paper does not provide multiple seeds or a significance test across seeds. Given that REINFORCE with T=80 steps is high-variance, single-seed reporting is insufficient to support 'matches' or the +20.4-point gain cell, especially since the ASSDM severity value at 10x (0.616) appears non-monotonic relative to its 20x value (0.680). Please report multiple seeds or an explicit test that the differences are outside noise.","section":"Table 1, §4.1, §4.2"},{"comment":"The ASSDM baseline is retrained from scratch under 'our exact data conditions,' but the paper gives no indication of the hyperparameter search or number of restarts for the baseline, and its non-monotonic AUC curve across accelerations (Table 1, severity: 0.680 at 20x, 0.616 at 10x, 0.695 at 6x) suggests the comparison may not be a stable estimate of the baseline's performance. Because the paper's 'largest single-cell gain' is the difference to this unstable value, the magnitude of the claimed improvement over ASSDM is not yet well established. Please provide baseline training details, multiple seeds for the baseline, or a sensitivity analysis.","section":"§4.1 'Acceleration augmentation', Table 1"}],"minor_comments":[{"comment":"The abstract and headline results do not carry the single-seed and single-anatomy caveat that the conclusion acknowledges. For a medical-imaging claim, at least a sentence in the abstract limiting the scope would help readers calibrate the claim.","section":"Abstract, §5"},{"comment":"No code or learned masks are released, and Figure 2 shows only qualitative masks. Exact coordinates or a release of the trained policy would be needed for reproduction; please state availability.","section":"§4.1, Reproducibility"},{"comment":"The notation CE_t does not indicate which classifier is being used for the disease and severity rewards. The text says severity CE is 'set to zero on disease-negative samples,' but the equation displays both rewards without this caveat; please make this explicit in the formula or its surrounding text.","section":"§3.5, Eq. (3)"},{"comment":"In several Cartilage Severity cells, Ours-Cartesian and Ours-hierarchical report identical values (e.g., 0.625, 0.641, 0.648 at 20x, 10x, 8x). Please confirm that the action space has no effect there or add a sentence explaining the overlap.","section":"Table 2"},{"comment":"The cosine curriculum and the low-frequency foundation are motivated qualitatively, but no ablation quantifies their contribution. The conclusion acknowledges this; a brief ablation or a clear statement that this is deferred would strengthen the paper.","section":"§3.2, §5"}],"recommendation":"major_revision","confidential_remarks":"The circularity between reward and evaluation is the main technical concern; I would require an independent-classifier or joint-fine-tuning evaluation before publication. The single-seed issue compounds this. The paper presents a useful idea and clean ablations, but the headline claims are not yet supported at the current level of evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one if you care about task-driven acquisition. The framework is sensible: a cosine-annealed curriculum that keeps a low-frequency disk always sampled, a Mamba policy that selects individual high-frequency points, and a reward based on per-step cross-entropy reduction of a frozen classifier. The action-space ablation (rows vs points) is the right kind of control and shows a real, if modest, benefit on cartilage diagnosis. The writing is clear and the limitations are listed.\n\nWhat's new: the combination of hierarchical frequency action space, curriculum, and Mamba policy is not in the prior work I know. The use of class-weighted CE reduction as a reward is standard in active learning but applied here in a sensible way.\n\nWhere it's soft: the single seed is the biggest problem. The bootstrap intervals are reported, but they are not a substitute for multiple runs. The fact that ASSDM's ACL severity AUC is 0.680 at 20x and 0.616 at 10x makes the headline +20.4 point gain fragile; that baseline looks poorly behaved. The reward/eval loop using the same frozen classifier is real but not a fatal flaw — it is internally consistent, and the paper doesn't claim the policy transfers to a different classifier. Still, the abstract says \"diagnostic accuracy,\" which implies more than one frozen model. A second classifier or a human read would strengthen the claim substantially. The ablation does not separate the cosine schedule from the low-frequency disk, and the code and masks are not released — that hurts reproducibility.\n\nThe paper is worth engaging with. It is not a breakthrough, but it is a coherent step forward in a practical direction. I would send it to peer review, with the conditions the reader listed: multi-seed results, code/masks, an independent classifier (or at least a clear statement that results are tied to this classifier), and a random sampling baseline. The action-space comparison with a shared backbone is a good control and should stay.","headline":"A clearly-written task-driven active sampling paper with a genuinely new action-space design; the evidence supports a conditional acceptance, not the headline as stated.","tokens_in":8175,"tokens_out":2492,"would_cite":false,"duration_ms":23368,"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":"This paper claims that an active MRI sampling strategy that treats low and high spatial frequencies differently can match fully-sampled diagnostic accuracy on ACL injury detection at 4x-10x acceleration, while also improving severity gradin","keywords":["active sampling","accelerated MRI","k-space sampling","Mamba","task-driven acquisition","diagnostic accuracy","ACL injury","reinforcement learning"],"falsifier":"Train a second, independently initialized classifier on the same data, run HieraSample's learned masks through it, and check whether its AUC on ACL diagnosis and severity stays within the reported ±0.003/±0.015 bootstrap intervals; if the advantage over the Cartesian baseline vanishes, the reward signal is overfit to the original classifier's loss surface rather than reflecting diagnostic content.","tokens_in":7161,"feed_emoji":"🧲","tokens_out":6974,"duration_ms":60708,"temperature":0.7,"pith_summary":"This paper aims to show that active k-space sampling for accelerated MRI can be guided by the clinical task itself rather than by generic image-quality metrics. It proposes a sampling policy that always keeps a fully-sampled low-frequency region and uses a state-space (Mamba) model to pick individual high-frequency coordinates one by one, guided by how much each new point reduces the classification loss on ACL injury and severity. On a large public knee MRI benchmark, the authors report that this policy matches the fully-sampled classifier's ROC AUC on ACL diagnosis at 4x-10x acceleration (0.939 AUC at 10x, 8x, and 6x) and improves severity grading by as much as 20.4 AUC points over a Cartesian-row baseline. If true, this would mean diagnostic accuracy does not have to be sacrificed for the 4-10x scan-time reductions central to clinical MRI.","feed_headline":"Active MRI sampling matches fully-sampled ACL diagnosis at 10x","feed_subtitle":"A frequency-aware policy preserves ACL diagnostic AUC from 4x to 10x and lifts severity grading by up to 20 points.","key_machinery":"The load-bearing mechanism is the frequency hierarchy embedded in the action space. At every step the mask contains a fully-sampled low-frequency disk (the smallest disk around DC), fixed before any policy decision, while a Mamba state-space policy selects only high-frequency coordinates, one point at a time, up to a per-step budget set by a cosine-annealed acceleration schedule from 20x to 4x. The policy is trained by REINFORCE with a reward equal to the per-sample reduction in class-weighted cross-entropy of a frozen Mamba-based classifier, so positive reward means the newly added point made the correct classification more confident. The paper describes normalization matching between class","core_discovery":"The central discovery claim is that a frequency-hierarchical action space - a preserved low-frequency disk plus point-level selection of high-frequency k-space coordinates - combined with a reward that is the per-step reduction in class-weighted cross-entropy, drives an active sampler to near-oracle diagnostic performance. The method's cosine-annealed curriculum moves acceleration from 20x to 4x over 80 steps, and the Mamba policy chooses individual coordinates from features of dual disease and severity classifiers. The paper reports that on ACL diagnosis, the learned masks achieve a ROC AUC of 0.939 at 10x, 8x, and 6x acceleration, identical to the fully-sampled oracle within bootstrap erro","pith_inferences":["Because the reward is tied to one frozen classifier, the 'matches oracle' claim may not transfer to a different classifier, a different reconstruction, or a radiologist; an independent classifier or human-reader study would test this directly.","The paper leaves the cosine schedule and low-frequency foundation as a combined design; a factorial ablation could reveal whether the gains come from the schedule, the fixed low-frequency disk, or their interaction.","If the hierarchy assumption is right, analogous point-based active sampling might help other diagnostic tasks whose features live at high spatial frequencies (e.g., microcalcifications in mammography), but that is an extrapolation beyond the paper's evidence.","Under the reported 0.4s per-slice rollout cost, the approach could support real-time adaptive scanning; whether the policy's choices remain robust to motion and other artifacts is a separate, untested question."],"forward_implications":["ACL diagnosis ROC AUC reaches 0.939 at 10x, 8x, and 6x acceleration, matching the fully-sampled classifier within bootstrap error.","ACL severity AUC rises from 0.616 (Cartesian baseline at 10x) to 0.820, a 20.4-point gain.","Cartilage diagnosis also improves, with the point-based action space gaining up to 0.015 AUC over rows at 20x, exceeding the reported confidence interval.","The per-step cross-entropy reduction reward yields positive signal across an 80-step rollout, supporting its use as a task-driven acquisition objective.","A rollout costs about 0.4 seconds per slice on a single GPU, indicating the policy can run at scan time."],"fun_headline_variants":["Frequency-hierarchical sampling matches full MRI at 10x","Active k-space sampling: ACL diagnosis matches full scan at 10x","HieraSample: MRI at 10x matches full-sampled ACL diagnosis","Point-wise k-space selection preserves ACL accuracy at 10x","MRI active sampling hits oracle-level ACL AUC up to 10x"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The paper assumes that reducing the per-step cross-entropy of a single frozen classifier is a faithful reward for improving true diagnostic accuracy, and that the same classifier's features are sufficient for the policy to choose clinically useful high-frequency points.","fun_headline_variants_meta":{"raw":{"variants":["Frequency-hierarchical sampling matches full MRI at 10x","Active k-space sampling: ACL diagnosis matches full scan at 10x","HieraSample: MRI at 10x matches full-sampled ACL diagnosis","Point-wise k-space selection preserves ACL accuracy at 10x","MRI active sampling hits oracle-level ACL AUC up to 10x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00017,"raw_usage":{"total_tokens":1100,"prompt_tokens":735,"completion_tokens":365,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":479,"completion_tokens_details":{"reasoning_tokens":272}},"tokens_in":479,"tokens_out":365,"duration_ms":3912,"temperature":1.0,"reasoning_tokens":272,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T11:44:06.840710+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a second, independently initialized classifier on the same data, run HieraSample's learned masks through it, and check whether its AUC on ACL diagnosis and severity stays within the reported ±0.003/±0.015 bootstrap intervals; if the advantage over the Cartesian baseline vanishes, the reward signal is overfit to the original classifier's loss surface rather than reflecting diagnostic content.","supporting_citations":[],"review_version":1}