{"id":"dca91d53-f58d-491b-932a-eb1ca1f0f365","arxiv_id":"2506.17412","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A recurrent vision-Mamba model with longitudinal asymmetry tracking improves 1-5 year breast cancer risk prediction on CSAW-CC, especially in high-density cases.","lead":"This paper tests a new deep-learning model that reads five years of screening mammograms to predict a woman's chance of developing breast cancer over the next one to five years. The model combines a recurrent memory network with a measure of asymmetry between the left and right breast, and the authors report better scores than existing models, especially for dense breasts and later years.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported improvements over LoMaR hinge on an unverified CSAW-CC split and temporal-filtering protocol delegated to reference [10]; if same-patient images leak or post-diagnosis exams are used as inputs, all headline C-index/AUC gains are inflated.","rationale":"The reader's weakest_assumption identifies the unvalidated LoMaR preprocessing and split protocol as the load-bearing premise for every numerical claim. My stress-test concurs: the paper never describes the train/test split or the temporal filtering of exams relative to diagnosis dates, and it simply defers to reference [10]. Without this information, the reported C-index and ROCAUC improvements cannot be interpreted, because both leaky splits and post-diagnosis inputs would inflate the metrics for a longitudinal model. This is the most fundamental concern because it affects all of Table 1, Table 2, and the headline claims, whereas the other issues (impossible CIs, p=0.061, loss mismatch, frozen Mirai encoder) are either secondary or would become clearer once the evaluation protocol is fixed. The verdict should remain CONDITIONAL: the authors must provide the split and filtering details, correct the table errors, and rerun significance tests. If the protocol turns out to be invalid, the verdict would move to REJECT; if it is valid and the improvements persist, the paper could be accepted. No change to the reader's verdict is therefore needed at this stage.","tokens_in":8288,"tokens_out":6111,"duration_ms":67396,"concrete_test":"Run the provided GitHub code on CSAW-CC and extract the exact patient IDs in the train/test splits and the exam dates used as inputs. Verify (1) no patient appears in both splits, and (2) every input mammogram for a cancer case is dated strictly before the diagnosis date. If either condition fails, re-run all experiments with a clean patient-level split and temporal filtering, then recompute Table 1 and Table 2. If the VMRA-MaR versus LoMaR differences at 4–5 years and in the high-density 1-year AUC shrink to within the reported confidence intervals, the central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central numerical claim—VMRA-MaR surpasses LoMaR, especially at 4–5 years and in high-density subgroups—depends entirely on a valid and fair evaluation on CSAW-CC. Section 3.1 states that the authors 'adhere to the preprocessing protocol outlined in LoMaR [10]' and delegates split details to that reference, but no patient-level split, date filtering, or exclusion of post-diagnosis exams is described in the manuscript. If same-patient mammograms appear in both train and test, or if exams acquired after the cancer diagnosis are included in the input sequence for cases, the C-index and ROCAUC for VMRA-MaR are inflated and the comparison to LoMaR is meaningless. This is not a minor detail: the model uses five yearly screenings per patient, so temporal leakage is easy to introduce inadvertently. The manuscript also reports borderline significance (p=0.061) for the overall comparison, and Table 1 contains internally inconsistent confidence intervals (e.g., LoMaR 3-year ROCAUC 0.83 with 95% CI 0.85–0.86), undermining confidence that even the reported numbers are correctly computed. A second confound is that VMRA-MaR's image encoder is initialized with frozen Mirai weights (Section 3.2); if LoMaR does not use the same pretrained encoder, part of the gain may be attributable to the encoder rather than the proposed temporal/asymmetry modules.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes VMRA-MaR, a recurrent framework for 1–5 year breast cancer risk prediction from longitudinal mammograms. It combines a VMRNN (Vision Mamba RNN) temporal encoder with a Spatial Asymmetry Detector and a Longitudinal Asymmetry Tracker, and feeds the fused representation into an additive hazard layer. The method is evaluated on the CSAW-CC dataset and compared with OncoNet, Mirai, AsymMirai, and LoMaR, with claims of improved performance especially at 4–5 year follow-up and in high-density subgroups.","tokens_in":8608,"tokens_out":5754,"duration_ms":56930,"significance":"If the reported evaluation is valid, the paper would make a useful empirical contribution: it demonstrates a recurrent state-space temporal model as an alternative to Transformer-based longitudinal baselines for mammography risk prediction, and it extends asymmetry analysis from single time steps to a longitudinal setting. The authors also release code and build on public datasets and pretrained weights, which aids reproducibility. However, because the central comparisons rest on an underspecified data split, internally inconsistent confidence intervals, and a borderline significance level, the significance is currently conditional on major revisions to the evaluation and reporting.","major_comments":[{"comment":"The CSAW-CC split and temporal-filtering protocol are not described in the manuscript. Section 3.1 states only that the authors 'adhere to the preprocessing protocol outlined in LoMaR [10]' and delegate the details to that reference. Since the model consumes five yearly screenings per patient, this is load-bearing: if same-patient images appear in both training and test sets, or if exams acquired after diagnosis are included in the input sequences, the reported C-index and ROCAUC gains over LoMaR would be inflated. The authors must specify the patient-level split, the exclusion of post-diagnosis exams, the matching of controls, and the exact follow-up definition, or the numerical claims cannot be interpreted.","section":"§3.1"},{"comment":"Many confidence intervals in Table 1 do not contain their reported point estimates; for example, LoMaR 3-year ROCAUC is 0.83 with 95% CI 0.85–0.86, 4-year is 0.82 with 0.84–0.86, and 5-year is 0.83 with 0.85–0.87, and similar inconsistencies appear in other rows. This indicates an error in confidence-interval computation or transcription. Because every comparison in the paper is expressed through these numbers, the table must be corrected before any conclusion about relative performance can be drawn. In addition, Table 2 reports subgroup ROCAUC values without any confidence intervals, which is especially problematic for the headline high-density 1-year result (0.97 vs 0.88 for LoMaR).","section":"Table 1"},{"comment":"The central claim that VMRA-MaR surpasses LoMaR, particularly at 4–5 years, is not consistently supported by the reported numbers. The overall comparison shows p=0.061 (borderline), and at 5 years VMR_MaR achieves ROCAUC 0.86 (0.84–0.86), numerically higher than VMRA-MaR's 0.84 (0.82–0.86). The text in §3.3 attributes long-horizon gains to combining VMR and Asym, which contradicts the table. The authors should report confidence intervals for the differences, specify the statistical test used for p, and provide a corrected interpretation of which module drives long-horizon performance.","section":"§3.3, Table 1"},{"comment":"The training objective is described inconsistently. Section 2.4 states that the additive hazard layer is optimized 'via log-likelihood maximization on the observed screening follow-up data,' while Section 3.2 says the model is trained with a 'weighted cross-entropy loss with class weights derived from the training set distribution, following LoMaR.' These are different objectives, and the AHL in Eqs. (7)–(8) estimates cumulative risk probabilities, not necessarily a categorical cross-entropy target. The authors need to state which loss is actually minimized, how the follow-up years are converted into labels, and how class weights are applied.","section":"§2.4 and §3.2"},{"comment":"The image encoder is initialized with Mirai weights and kept frozen during training, and VMRNN blocks are initialized from pretrained VMRNN weights. For a fair comparison, the manuscript must specify whether LoMaR and the other baselines use the same pretrained encoder and the same input views and risk-factor variables; otherwise the reported gains may be attributable to the encoder initialization rather than to the proposed temporal and asymmetry modules. The ablation 'VMR_MaR' helps isolate the asymmetry module only if the encoder is identical across the compared models.","section":"§3.2"}],"minor_comments":[{"comment":"Figure 1 contains multiple typos, including 'Spilt', 'ImgEmbddins', 'Embddins-5Y', and 'Aggerator'; these should be corrected for clarity.","section":"Figure 1"},{"comment":"The text refers to 'CSAW-CCS' in the density-stratification description; this should be 'CSAW-CC'.","section":"§3.4"},{"comment":"The phrase 'A baseline risk, B(eR), is computed via a Linear k(eR)' appears to contain a notation or formatting error; please clarify the intended expression.","section":"§2.4"},{"comment":"The 40% displacement threshold in the LAT and the choice of average fusion over concatenation are described without justification or sensitivity analysis; reporting experiments across threshold values would strengthen the paper.","section":"§2.3"},{"comment":"The p-value of 0.061 is reported without specifying the statistical test used or whether any multiple-comparison correction was applied; this information should be added.","section":"§3.3"},{"comment":"The caption states that five representative subjects are shown, then lists 'two high-, one medium-, and one low-density'; this sums to four subjects and should be corrected.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is best suited to a medical-imaging venue that requires strong empirical validation. The main risk is evaluation validity on CSAW-CC: if the split and temporal-filtering protocol cannot be reconstructed from LoMaR, the comparison would be uninterpretable. I would ask for the full protocol details, corrected confidence intervals, and a clarified loss function before any acceptance decision. The paper's novelty is reasonable but not exceptional; the value lies in the empirical comparison, which currently needs substantial repair."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new part here is real but modest: VMRNN's recurrent state-space encoding has not been applied to longitudinal mammography before, and extending AsymMirai's single-time-point asymmetry into a temporal tracker is a reasonable idea. The ablations (VMR_MaR vs VMRA_MaR) suggest the asymmetry module does something, and releasing code is a point in the authors' favor. The high-density 1-year AUC of 0.97 is eye-catching, but I would not trust it yet.\n\nThe soft spots are concentrated in the evidence, not the architecture. Table 1 contains impossible confidence intervals: LoMaR's 3-year ROCAUC is 0.83 with 95% CI 0.85–0.86, and the 5-year row is similarly broken. That alone says the numbers were not checked. The main comparison against LoMaR has p=0.061, which the authors disclose honestly, but they still write \"notable improvements\" and \"superior performance\" in the abstract and conclusion. Those words are not supported. The Section 2.4 log-likelihood vs Section 3.2 weighted cross-entropy mismatch is probably a wording issue, but it needs fixing.\n\nThe bigger problem is the evaluation protocol. The paper says it adheres to LoMaR's preprocessing and delegates the split to reference [10], with no patient-level split, no temporal filtering, and no exclusion of post-diagnosis exams described. If same-patient images leak across train/test, or if exams after diagnosis are used as inputs for cases, every C-index and AUC is inflated. LoMaR's protocol may be public and fine, but the authors have to show they followed it, not just cite it. The frozen Mirai encoder is a second confound: if LoMaR does not use the same pretrained encoder, part of the gain may be the encoder, not the VMRNN or asymmetry modules.\n\nThese are fixable problems. The architecture is plausible and the topic matters. I would send this to peer review, but only after the authors fix the table, describe the split and temporal filtering explicitly, add confidence intervals to the density subgroups, and clarify the loss function. As it stands, I would not cite it in my own work yet, but I would bring it to a reading group as a case study in how evaluation details, not model design, often decide whether a medical imaging paper holds up.","headline":"A sensible architecture paper whose headline gains rest on an under-described evaluation protocol and statistics that do not yet support the claims.","tokens_in":9104,"tokens_out":1844,"would_cite":false,"duration_ms":20733,"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":"VMRA-MaR claims that a recurrent Vision Mamba RNN combined with bilateral asymmetry tracking improves longitudinal breast cancer risk prediction over the LoMaR transformer baseline, particularly at 4-5 year follow-up horizons and in…","keywords":["breast cancer risk prediction","longitudinal mammograms","Vision Mamba RNN","state-space model","bilateral asymmetry","CSAW-CC dataset","temporal modeling"],"falsifier":"Re-run both VMRA-MaR and LoMaR under the same protocol with every exam dated after the cancer diagnosis excluded and all images from one woman kept in only one split, then compare the four- and five-year C-index and the high-density one-year ROCAUC; if the margins shrink toward the borderline p-value of 0.061, the claimed advantage is not established.","tokens_in":8103,"feed_emoji":"🩺","tokens_out":5590,"duration_ms":54960,"temperature":0.7,"pith_summary":"The paper sets out to establish that longitudinal breast cancer risk prediction improves when temporal modeling is recurrent and asymmetry-aware rather than transformer-based with positional embeddings. It proposes VMRA-MaR, which processes five yearly screening mammograms through a Vision Mamba RNN (VMRNN) with LSTM-like memory and an asymmetry module that tracks persistent left-right differences. The central empirical claim is that this combination outperforms the previous state-of-the-art LoMaR at the four- and five-year horizons, reaching a C-index of 0.82 with ROCAUC of 0.84, and shows a clear advantage in high-density breasts, with a one-year ROCAUC of 0.97 versus 0.88 for LoMaR. A sympathetic reader would care because longer-horizon and dense-breast predictions are exactly the clinically challenging cases where existing risk models are weakest.","feed_headline":"Recurrent vision model lifts long-term breast cancer risk scores","feed_subtitle":"At years 4-5, C-index 0.82 and ROCAUC 0.84; in high-density breasts, 1-year ROCAUC reaches 0.97.","key_machinery":"The central mechanism is the VMRNN block: a recurrent cell that updates a hidden state and cell state at each yearly screening visit using a gating signal from a Vision State Space (Mamba-style) module, with downsampling, upsampling, and a reconstruction layer for multi-scale representation. It is paired with a Spatial Asymmetry Detector (SAD) that aligns left and right views and finds the maximum local feature difference, and a Longitudinal Asymmetry Tracker (LAT) that follows that maximum's location across visits, upweighting the asymmetry score when the abnormality persists. The Additive Hazard Layer combines the temporal embedding with the asymmetry score into monotone, non-negative cumulative risk estimates for years one through five.","core_discovery":"On its own terms, the paper claims that longitudinal breast cancer risk prediction improves when temporal modeling is recurrent and asymmetry-aware rather than transformer-based with positional embeddings. The proposed VMRA-MaR processes five yearly screening visits, each with four mammographic views, through a frozen Mirai image encoder; the VMRNN block carries an LSTM-like cell state gated by a Vision State Space module, and the asymmetry module upweights regions whose contralateral difference persists across visits. As reported, this yields a C-index of 0.82 and ROCAUC of 0.84 at years four and five, beating LoMaR, and a one-year ROCAUC of 0.97 in the high-density subgroup. The paper also reports that averaging asymmetry scores over time outperforms concatenating them, and that Grad-CAM saliency maps place more weight near expert-annotated lesions when both temporal and asymmetry cues are present.","pith_inferences":["Editorial inference: If these gains survive on external screening cohorts, screening programs could use the model to tailor intervals by breast density rather than age alone, since the high-density one-year margin is the strongest reported result.","Editorial inference: The LAT's persistent-abnormality threshold (40% of the window size) is a free parameter; a natural extension is to test whether varying it changes the high-density advantage, which the paper does not report.","Editorial inference: Because the asymmetry module operates on bilateral feature maps, the same SAD plus LAT design could transfer to other paired-organ imaging settings, but the paper only validates it on mammograms."],"forward_implications":["Long-horizon risk prediction improves: the model reaches a C-index of 0.82 with ROCAUC of 0.84 at both year 4 and year 5, where the LoMaR baseline scores lower.","High-density breasts show the largest gain: the one-year ROCAUC of 0.97 versus 0.88 for LoMaR indicates better risk separation in the subgroup where mammography is most difficult.","Recurrent temporal encoding contributes independently of asymmetry: the VMRNN-only variant also improves some horizons, such as a five-year ROCAUC of 0.86 versus 0.83 for LoMaR.","Asymmetry tracking helps early horizons: VMRA-MaR reaches one-year ROCAUC of 0.94 and two-year ROCAUC of 0.88, compared with 0.92 and 0.83 for LoMaR.","Averaging asymmetry scores across time is the better fusion strategy, suggesting that persistent rather than one-time asymmetry carries the predictive signal."],"supporting_citations":[{"why":"Defines the LoMaR baseline and the preprocessing protocol that the paper follows, so every numerical comparison inherits this protocol.","marker":"[10]"},{"why":"Provides the asymmetry insight and the left-right alignment and difference computation that the Spatial Asymmetry Detector builds on.","marker":"[4]"},{"why":"Supplies the VMRNN block, its recurrent cell equations, and the pretrained weights used to initialize temporal modeling.","marker":"[18]"},{"why":"Supplies the frozen Mirai image encoder that produces the embeddings fed into the VMRNN and asymmetry modules.","marker":"[23]"},{"why":"Provides the CSAW-CC dataset with longitudinal mammograms, exam dates, and tumor annotations used for training and evaluation.","marker":"[17]"},{"why":"Defines the additive hazard layer that turns the temporal and asymmetry embeddings into monotone five-year risk estimates.","marker":"[11]"},{"why":"Supplies the Mamba-style state-space directional module (S6) used inside the VMRNN cell's gating stream.","marker":"[7]"},{"why":"Grounds the density subgroup stratification in absolute dense area rather than percentage density.","marker":"[16]"}],"fun_headline_variants":["Recurrent vision model boosts long-term breast cancer risk","Asymmetry-aware temporal model improves 4-5 year risk","Vision Mamba RNN lifts breast cancer risk prediction","State-space memory and asymmetry enhance risk scores","Longitudinal model excels at high-density breast risk"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The numerical gains over LoMaR stand on the validity of LoMaR's preprocessing and train/test split, which the paper delegates to reference [10] and does not describe; if exam dates after diagnosis are not filtered or the same woman's images appear in both train and test, the reported C-index and ROCAUC improvements could be inflated.","fun_headline_variants_meta":{"raw":{"variants":["Recurrent vision model boosts long-term breast cancer risk","Asymmetry-aware temporal model improves 4-5 year risk","Vision Mamba RNN lifts breast cancer risk prediction","State-space memory and asymmetry enhance risk scores","Longitudinal model excels at high-density breast risk"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000236,"raw_usage":{"total_tokens":1529,"prompt_tokens":997,"completion_tokens":532,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":456}},"tokens_in":613,"tokens_out":532,"duration_ms":6166,"temperature":1.0,"reasoning_tokens":456,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:30:51.449478+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run both VMRA-MaR and LoMaR under the same protocol with every exam dated after the cancer diagnosis excluded and all images from one woman kept in only one split, then compare the four- and five-year C-index and the high-density one-year ROCAUC; if the margins shrink toward the borderline p-value of 0.061, the claimed advantage is not established.","supporting_citations":[{"cited_title":"In: Linguraru, M.G., Dou, Q., Feragen, A., Giannarou, S., Glocker, B., Lekadir, K., Schnabel, J.A","cited_arxiv_id":null,"evidence_quote":"Defines the LoMaR baseline and the preprocessing protocol that the paper follows, so every numerical comparison inherits this protocol."},{"cited_title":"Radiology310(3), e232780 (2024)","cited_arxiv_id":null,"evidence_quote":"Provides the asymmetry insight and the left-right alignment and difference computation that the Spatial Asymmetry Detector builds on."},{"cited_title":"Science Translational Medicine13(578), eaba4373 (2021)","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen Mirai image encoder that produces the embeddings fed into the VMRNN and asymmetry modules."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the CSAW-CC dataset with longitudinal mammograms, exam dates, and tumor annotations used for training and evaluation."},{"cited_title":"Breast Cancer Research12, 1–7 (2010)","cited_arxiv_id":null,"evidence_quote":"Grounds the density subgroup stratification in absolute dense area rather than percentage density."}],"review_version":1}