{"id":"d2213623-4303-47f0-853b-ba6fdd766398","arxiv_id":"2505.11832","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Auto-RMP, an autoregressive VQGAN plus causal transformer model, predicts future 4D CT phases from prior phases and reports higher lung and heart motion accuracy than DAM and DiffuseRT on public and private datasets.","lead":"Researchers trained an autoregressive neural network to predict future breathing-phase CT images from a patient's earlier 4D CT frames, and report it beats two published motion-prediction models on lung and heart overlap scores. The method could support tighter radiation target margins in radiotherapy if the results can be independently reproduced.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported superiority of Auto-RMP rests on an unspecified train/test split; on a longitudinal dataset, intra-patient leakage could explain the large gains, so the performance claim is not yet established.","rationale":"I agree with the reader's conditional verdict and with the identified weakest assumption. The strongest claim is the quantitative superiority in Table 1, and the load-bearing condition is that the evaluation isolates generalization to unseen patients or phases. The paper never specifies the train/test split for either dataset, and the longitudinal structure of the Hugo dataset makes same-patient leakage a live possibility rather than a hypothetical. The proposed test—patient-level cross-validation with baselines run under identical inputs—would settle whether the reported gap persists. If the gap disappears, the contribution reduces to interpolation within already-seen patients, which is much weaker than the claimed patient-specific generalization. I do not see an internal contradiction in the method itself; the issue is under-specification of the evaluation, so the appropriate verdict remains conditional acceptance pending the missing details.","tokens_in":8853,"tokens_out":4981,"duration_ms":57256,"concrete_test":"Ask the authors to release the exact patient-level train/test split and rerun Table 1 with leave-one-patient-out cross-validation on the public Hugo dataset (20 patients, 80 scans), reporting per-patient mean and standard deviation for Auto-RMP, DAM, and DiffuseRT under identical input modalities. If the public IoU/DSC gains shrink to within error bars or drop by several points, the outperformance claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim (Table 1, Section 3.2) depends on an evaluation protocol that is never described: the paper does not state how the public or private 4D CT scans were divided into training and testing sets. This matters because the public Hugo dataset is longitudinal (20 patients, 80 4D CT scans, Section 3.1), so repeated scans from the same patient can easily appear in both training and testing. If that happens, the autoregressive transformer can memorize patient-specific anatomy rather than learn a generalizable motion model, which would directly inflate the reported 90.75% IoU / 95.15% DSC for lung and 88.43% IoU / 93.85% DSC for heart. The paper also omits how the two baseline methods (DAM, DiffuseRT) were configured, whether they received the same CT-plus-mask input, and how metrics were pooled across patients and phases. Without this information, the reported improvements cannot be checked. A closely related risk is that the same TotalSegmentator masks are used both as auxiliary input and as ground truth for evaluation, so part of the apparent accuracy may reflect propagation of the input mask rather than true prediction of future motion. This is not a cosmetic omission: the paper's strongest claim is precisely that Auto-RMP significantly outperforms existing methods, and that claim currently rests on numbers that could be produced by data leakage or an unfair comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces Auto-RMP, an autoregressive model for predicting future 4D CT phases in lung and heart radiotherapy motion. After segmenting organs with TotalSegmentator/nnUNet, each CT phase is tokenized by a VQGAN into discrete tokens, and a LLaMA-style causal transformer predicts the next phase's tokens conditioned on previous phases. The method is evaluated on the public Hugo dataset (20 patients, 80 scans) and a private 50-patient dataset against DAM and DiffuseRT, reporting higher IoU/DSC and lower surface distances for next-phase and multi-phase prediction. An ablation compares CT-only, mask-only, and CT+mask inputs.","tokens_in":9140,"tokens_out":5979,"duration_ms":56872,"significance":"The core idea, casting organ motion as autoregressive sequence generation, is a clean reformulation that avoids DVF dependence and enables multi-phase roll-out. If the reported numbers survive a correctly specified evaluation, this would be a practically relevant advance for pre-treatment motion management. The paper motivates the work well, and the token-level formulation is simple enough to reproduce. The main contribution is currently empirical, and the empirical claims are not yet verifiable because evaluation protocol details are missing.","major_comments":[{"comment":"The manuscript never states how the 80 public and 50 private 4D CT scans were split into training, validation, and test sets. The public Hugo dataset is longitudinal (20 patients, 80 scans), so if scans from the same patient appear in both training and testing, the reported patient-specific accuracy may reflect memorization of that patient's anatomy rather than generalization. Please report patient-level splits, the number of patients and scans in each split, and per-patient test metrics (or confidence intervals) to rule out leakage.","section":"§3.1 Dataset / §3.2 Table 1"},{"comment":"The same TotalSegmentator nnUNet model is described as providing ground-truth labels for organ shape and position, and those masks are also concatenated with CT as input to the model. This creates a circular evaluation: high IoU/DSC may partly measure how well the model propagates the input mask rather than true motion prediction. Please clarify whether ground-truth masks for evaluation are independent of the input masks (e.g., manual contours or a different tool), or quantify the sensitivity of the reported metrics to segmentation noise.","section":"§2.1 Pre-processing / §3.1 Evaluation"},{"comment":"No configuration details are given for the DAM and DiffuseRT baselines: architecture, hyperparameters, training data, input format (CT only versus CT+mask), context length, or compute budget. Without these, 'significantly outperforms' cannot be verified; the comparison may be unfair. Please specify baseline setups, report the same splits and metrics for all methods, and make code and checkpoints available.","section":"§3.2 Table 1"},{"comment":"Long-term motion prediction is claimed as a key advantage, but the evidence is a qualitative figure and a single sentence stating that Auto-RMP maintains over 85% DSC in the final predicted phase. No per-phase quantitative table, no error bars, and no baseline comparison are provided. Please report per-phase IoU/DSC/SD/HD for Auto-RMP and all baselines in the five-phase roll-out setting.","section":"§3.2 Long-term prediction"},{"comment":"All reported numbers are point estimates without standard deviations, confidence intervals, or statistical tests across patients and phases. The caption of Table 1 uses 'significantly outperforms,' which is not supported by any significance test. Please include per-patient variability and paired tests (e.g., Wilcoxon signed-rank) for the main comparisons.","section":"Table 1 / Table 2"}],"minor_comments":[{"comment":"The caption refers to 'NSD' values, but the table contains no NSD column; please correct this.","section":"Table 1 caption"},{"comment":"The sentence 'We adopt pre-trained parameters from Yutong et al. [1]' should cite Bai et al., matching reference [1].","section":"§2.2 Implementation Details"},{"comment":"The caption contains a grammatical error: 'We first arranges each phase' should be 'We first arrange each phase.'","section":"Figure 2 caption"},{"comment":"The ablation uses a 'Mixed' set (20 public + 20 private scans) but does not describe which private scans were selected or how they were split; please specify this.","section":"Table 2"},{"comment":"Equation (1) has typographical spacing issues in the product limit notation, and the notation for the final phase X_T should be made consistent throughout.","section":"Equation (1)"}],"recommendation":"major_revision","confidential_remarks":"The evaluation-leakage concern is the main risk to the central claim; I would make patient-level split reporting and baseline configuration a hard requirement of revision. The paper's citation list contains several self-citations to preprints that are not directly related to the method; this is not a basis for rejection but worth noting to the authors. The work is on-topic for a medical imaging or radiotherapy physics journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the core idea is genuinely useful: treating organ motion prediction as an autoregressive sequence problem over VQGAN tokens is a clean way to get multi-phase rollouts without precomputed DVFs. That is a real contribution, even though every component is borrowed. Second, the reported 8–10 point IoU gains over DAM and DiffuseRT are not yet credible, because the paper never tells you how the data were split into train and test.\n\nThe strength is the framing. The authors argue—correctly—that 4D CT is a temporal sequence, and that autoregressive modeling is a natural fit. The pipeline is standard (VQGAN tokenization, causal transformer, next-phase prediction), but the application is new and the logic is coherent. They also run a useful ablation on input modalities, showing that combining CT and masks helps. The limitations section honestly notes that the evaluation is single-session only, which is the right kind of humility.\n\nThe soft spot is the evaluation, and it is load-bearing. The public Hugo dataset is longitudinal: 20 patients, 80 scans, with multiple scans per patient. If the split is by scan rather than by patient, the model can memorize patient anatomy, which would directly inflate the numbers. The paper does not state the split. It also does not say how DAM and DiffuseRT were configured, whether they received the same inputs, or how metrics were pooled across phases. No error bars, no code. On top of that, the ground-truth masks come from the same TotalSegmentator model that provides auxiliary input, so part of the apparent accuracy may be mask propagation rather than true motion prediction. These are all fixable with additional reporting, but right now Table 1 is a claim without a verifiable protocol.\n\nWho gets value from this? Researchers working on radiotherapy motion management and medical image sequence modeling. The paper is worth a serious referee, but the current version should be returned for major revision: clarify the split, describe baseline setups, add error bars and code, and discuss the mask-input circularity explicitly. If the numbers hold up after that, this becomes a solid application paper. As it stands, treat the quantitative claim as conditional.","headline":"A sensible repackaging of autoregressive token prediction for 4D CT organ motion, but the headline accuracy numbers rest on an evaluation protocol that is not described.","tokens_in":9646,"tokens_out":1741,"would_cite":false,"duration_ms":19502,"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":"This paper claims that organ motion in radiotherapy can be predicted as an autoregressive sequence of CT phases, and that the resulting model beats published DVF- and diffusion-based baselines on lung and heart motion.","keywords":["autoregressive model","organ motion prediction","radiotherapy","4D CT","VQGAN tokenization","patient-specific motion","lung and heart motion prediction"],"falsifier":"Run leave-one-patient-out cross-validation on the public Hugo dataset, holding out all scans of each patient together, and compare Auto-RMP's lung and heart IoU and DSC against the same baselines; if the gap narrows to the level of DAM or DiffuseRT, the claim that the model learns patient-specific motion would be unsupported. A complementary check would be to re-score predictions using manually curated contours instead of the same nnUNet/TotalSegmentator masks used for training input.","tokens_in":8647,"feed_emoji":"🩻","tokens_out":6743,"duration_ms":64681,"temperature":0.7,"pith_summary":"This paper claims that organ motion seen in radiotherapy can be predicted by treating a patient's 4D CT scan as a sequence and asking an autoregressive model to generate the next phase, the way language models generate the next token. The payoff would be practical: if future lung and heart positions can be forecast from the patient's own pre-treatment scans, radiation margins could be tightened and healthy tissue spared. The authors build Auto-RMP, a VQGAN-plus-causal-transformer pipeline, and report that it outperforms a DVF-based method (DAM) and a diffusion method (DiffuseRT) on both a public lung-cancer dataset and a private clinical dataset, reaching 90.75% IoU and 95.15% DSC for lung and 88.43% IoU and 93.85% DSC for heart on the public data. The same autoregressive setup also rolls out five future phases while keeping DSC above 85% on the last predicted phase.","feed_headline":"Autoregressive model predicts lung and heart motion from CT phases","feed_subtitle":"Treating 4D CT phases like text tokens reaches ~91% IoU on lung motion and rolls out future phases.","key_machinery":"The load-bearing mechanism is the autoregressive factorization in Equation (1): the joint probability of all future phases is written as a product of next-phase conditional probabilities, so training and inference reduce to next-token prediction. Around that factorization, Auto-RMP wraps two components: VQGAN, which encodes each CT phase into a 16x16 grid of discrete tokens with a downsampling factor of 16 and a codebook size of 8192, and a unidirectional causal transformer, based on the LLaMA architecture with a 4096-token context, which predicts the token sequence of the next phase given all previous phase tokens; the VQGAN decoder then renders the predicted phase. As a result, no deformable image registration or DVF computation is required, and the model can feed its own predictions back into the context to generate multiple future phases.","core_discovery":"On the paper's own terms, the central claim is that organ motion prediction is an autoregressive process over CT phases. Given the sequence of prior phases $X_0, \\ldots, X_T$, the model maximizes the product of conditional likelihoods $p_\\theta(X'_t \\mid X_0, \\ldots, X_{t-1})$, so that at inference the patient's earlier scans serve as a prefix from which future phases are generated one at a time. Auto-RMP implements this by tokenizing each 256x256 CT phase, concatenated with its organ mask, using VQGAN into 256 discrete tokens drawn from an 8192-entry codebook, then predicting the next phase's tokens with a causal LLaMA-style transformer and decoding them back into a CT image. Because conditioning is on the patient's own phase history rather than on a single image or a precomputed deformation vector field, the paper argues that the model captures patient-specific periodic motion and can be rolled forward for several phases without the error collapse it reports for single-step baselines. The reported evidence is the next-phase and multi-phase accuracy on the public and private datasets.","pith_inferences":["Editorial inference: the decisive test the paper does not report is a strict per-patient split on the longitudinal public dataset; if accuracy holds under leave-one-patient-out evaluation, the patient-specific claim is much stronger than if scans of the same patient appear in both training and testing.","Editorial inference: because the same nnUNet/TotalSegmentator pipeline produces both the auxiliary masks and the evaluation ground truth, an independent set of manually curated contours would clarify how much of the reported IoU and DSC reflects true anatomical prediction rather than shared segmentation bias.","Editorial inference: the same token-sequence framing could transfer to other periodic or quasi-periodic anatomies, such as liver motion in abdominal radiotherapy, and to other image modalities like cone-beam CT where longitudinal training data are beginning to exist.","Editorial inference: a head-to-head comparison with diffusion models that are themselves conditioned on multiple prior phases would isolate the contribution of the autoregressive conditioning from the choice of generative backbone."],"forward_implications":["Pre-treatment planning could use the predicted motion envelope to define smaller radiation margins, reducing dose to healthy tissue around the tumor.","Motion prediction no longer depends on deformable image registration: the model learns motion directly from CT phase sequences, removing a costly and error-prone preprocessing step.","Because the model conditions on the patient's own phase history, its accuracy is expected to track that patient's breathing pattern rather than a population average.","Multi-phase rollout means the same model can supply a full breathing cycle of future positions, not just the next phase, matching the temporal horizon needed for gated or adaptive delivery."],"supporting_citations":[{"why":"supplies the public longitudinal 4D CT dataset (20 patients, 80 scans, 800 phases) used for the public evaluation.","marker":"[14]"},{"why":"is the deformation-based deep-learning baseline (DAM) whose reported scores Auto-RMP is compared with in Table 1.","marker":"[30]"},{"why":"is the diffusion-based baseline (DiffuseRT) whose reported scores Auto-RMP is compared with in Table 1.","marker":"[34]"},{"why":"supplies the sequential-modeling approach and the pre-trained VQGAN weights that Auto-RMP adopts for token-level prediction.","marker":"[1]"},{"why":"defines the VQGAN encoder-decoder and quantization codebook used to tokenize CT phases and decode predictions.","marker":"[10]"},{"why":"provides the nnUNet segmentation framework used to produce organ masks from CT images.","marker":"[15]"},{"why":"provides the TotalSegmentator pre-trained weights that generate the lung and heart masks used as inputs and evaluation labels.","marker":"[41]"},{"why":"supplies the LLaMA transformer architecture used as the causal autoregressive backbone.","marker":"[35]"}],"fun_headline_variants":["CT phases as tokens: autoregressive organ motion prediction","Patient-specific autoregression predicts future CT phases","Beyond PCA: autoregressive forecasting of lung and heart motion","Treating 4D CT like text improves motion prediction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper never states how the public or private 4D CT scans were divided into training and testing sets, so if scans from the same patient appear on both sides, the reported patient-specific accuracy could reflect memorized anatomy rather than generalization to a new patient.","fun_headline_variants_meta":{"raw":{"variants":["CT phases as tokens: autoregressive organ motion prediction","Patient-specific autoregression predicts future CT phases","Beyond PCA: autoregressive forecasting of lung and heart motion","Treating 4D CT like text improves motion prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000253,"raw_usage":{"total_tokens":1624,"prompt_tokens":1064,"completion_tokens":560,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":680,"completion_tokens_details":{"reasoning_tokens":497}},"tokens_in":680,"tokens_out":560,"duration_ms":5924,"temperature":1.0,"reasoning_tokens":497,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:46:43.672253+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run leave-one-patient-out cross-validation on the public Hugo dataset, holding out all scans of each patient together, and compare Auto-RMP's lung and heart IoU and DSC against the same baselines; if the gap narrows to the level of DAM or DiffuseRT, the claim that the model learns patient-specific motion would be unsupported. A complementary check would be to re-score predictions using manually curated contours instead of the same nnUNet/TotalSegmentator masks used for training input.","supporting_citations":[{"cited_title":"Medical physics44(2), 762–771 (2017) 3, 4, 6","cited_arxiv_id":null,"evidence_quote":"supplies the public longitudinal 4D CT dataset (20 patients, 80 scans, 800 phases) used for the public evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"is the deformation-based deep-learning baseline (DAM) whose reported scores Auto-RMP is compared with in Table 1."},{"cited_title":"Physics in Medicine & Bi- ology 69(15), 155016 (2024) 2, 3, 6, 7","cited_arxiv_id":null,"evidence_quote":"is the diffusion-based baseline (DiffuseRT) whose reported scores Auto-RMP is compared with in Table 1."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"supplies the sequential-modeling approach and the pre-trained VQGAN weights that Auto-RMP adopts for token-level prediction."},{"cited_title":"In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"defines the VQGAN encoder-decoder and quantization codebook used to tokenize CT phases and decode predictions."},{"cited_title":"Nature Methods 18(2), 203–211 (2021) 5","cited_arxiv_id":null,"evidence_quote":"provides the nnUNet segmentation framework used to produce organ masks from CT images."},{"cited_title":"Radiology: Artificial Intelligence 5(5) (2023) 5","cited_arxiv_id":null,"evidence_quote":"provides the TotalSegmentator pre-trained weights that generate the lung and heart masks used as inputs and evaluation labels."}],"review_version":1}