{"id":"805a845a-ae1e-4ed6-a6d0-db845bcbe8b5","arxiv_id":"2411.16314","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Motion fields from 4D lung CT are reduced to oriented histograms and embedded with UMAP, producing patient clusters that appear to reflect breathing pattern similarity.","lead":"This paper turns the 3D motion fields from lung 4D CT scans into compact direction histograms and embeds them in a 2D map, so patients with similar breathing motion cluster together. It is a feasibility study for using pre-treatment imaging to compare a new patient to previously treated similar patients in radiotherapy.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'meaningful clustering' claim rests on visual inspection and same-patient repeat scans, but neither controls for anatomy, acquisition, or UMAP noise, so motion-specificity is not yet demonstrated.","rationale":"The reader's weakest_assumption focuses on information loss in the oriented histogram: localization is discarded and validation is visual. My concern is closely related but shifts the emphasis to confounded validation: even if the histogram retains motion information, the observed clustering could be driven by anatomy, acquisition protocol, or UMAP artifacts rather than by breathing similarity. The repeat-scan evidence is particularly vulnerable because patient identity is entangled with non-motion factors. This is a concrete, testable concern rather than a demonstration of fraud or error. I agree with the reader's conditional verdict: the method is coherent and the feasibility demonstration is plausible, but the central claim of meaningful clustering needs a quantitative control before it can be accepted as robust. I chose UNCHANGED because the reader's CONDITIONAL verdict already encodes this need; the direction-scrambled control would supply the missing evidence. I marked agreement as partial rather than full because the reader's stated weakest assumption is about information retention, whereas my attack is about whether the reported clustering actually reflects motion when non-motion confounds are present.","tokens_in":6117,"tokens_out":6583,"duration_ms":74989,"concrete_test":"Run the full pipeline on a direction-scrambled control: for every patient in the 4D-Lung set, keep the lung-mask eigenbasis and the per-voxel displacement magnitude r, but permute the (theta, phi) direction assigned to each voxel within each moving phase. This preserves anatomy, image acquisition, and amplitude statistics while destroying the spatial and directional structure of breathing motion. Compute the same-patient vs different-patient separation (e.g., AUROC of a patient-same classifier from pairwise UMAP distances, or silhouette score) for the real and scrambled embeddings. If the scrambled control matches the real AUROC, the repeat-scan clustering is not evidence for motion-specific embedding; if it does not, the directional motion signal is real.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the 2D UMAP embedding clusters 4D CT data sets by breathing-motion similarity. The evidence offered in Sec. 3 is (i) visual inspection of Fig. 2, where 'patients with vector fields with similar characteristics ... tend to form a cluster,' and (ii) the observation that repeat 4D CTs of the same patient in the 4D-Lung set mostly cluster together. Both are load-bearing because they are the only support for the statement in Sec. 4 that 'the embedding appears to be meaningful.' The problem is that neither observation isolates breathing motion from correlated nuisance structure. The pipeline intentionally retains patient-specific anatomy through the lung-mask eigenbasis used to define the spherical coordinate frame (Sec. 2.3, Eq. 1-2), and the 4D-Lung repeat scans share not only the patient but also the scanner, reconstruction protocol, and mask geometry. With n=71/33 and UMAP n_neighbors=5, apparent clusters can arise from sampling noise or from these confounds even if the histograms carried zero breathing information. The absence of any quantitative cluster-quality metric, baseline comparison, or stability analysis means the central claim currently is not distinguished from 'UMAP clusters some small dataset.' The paper is a plausible feasibility suggestion, but the meaningfulness of the embedding is not yet demonstrated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an unsupervised dimensionality-reduction pipeline for deformable image registration (DIR) vector fields from 4D CT lung data. Each vector field is transformed to a patient-specific lung basis, converted to spherical coordinates, and binned into a radius-weighted 8×16 oriented histogram per phase; the nine histograms of a 10-phase 4D CT are then compressed with a lightweight autoencoder and embedded in 2D with UMAP (n_neighbors=5, cosine metric, DensMAP). The authors demonstrate the approach on 71 in-house and 33 external 4D-Lung data sets, and argue from visual inspection of the UMAP plots, and from repeat 4D CTs of the same patient clustering together, that the embedding captures breathing-pattern similarity. They also state that the method is a general, computationally efficient way to reduce DIR vector fields for patient similarity analysis in radiotherapy.","tokens_in":6407,"tokens_out":3060,"duration_ms":30995,"significance":"If the central claim holds, the paper would provide a practical, fully unsupervised tool for comparing lung-motion patterns across patients using only planning 4D CT data, with potential applications in treatment-plan review, outcome analysis, and even hyperparameter selection for DIR (as claimed in the Learn2Reg context). The methodological idea of encoding directional motion statistics as oriented histograms is a reasonable and computationally attractive extension of 2D histogram-of-optical-flow methods to 3D registration fields, and the authors are right that this specific combination is not common in the radiotherapy literature. The paper is written clearly and the pipeline is straightforward to implement. However, the evidential basis for the main claim is currently qualitative: the paper relies on visual cluster inspection and on same-patient repeat scans, neither of which isolates breathing-motion similarity from anatomical or acquisition-related confounds. The absence of quantitative cluster-quality metrics, baseline comparisons, or stability analyses means the 'meaningful embedding' claim is not yet distinguished from clustering artifacts of a small dataset.","major_comments":[{"comment":"The central claim that clusters reflect breathing-pattern similarity is supported only by qualitative visual inspection ('patients with vector fields with similar characteristics ... tend to form a cluster') and by the observation that repeat 4D CTs of the same patient in the 4D-Lung set mostly cluster together. This is load-bearing because the abstract and the Discussion ('the embedding appears to be meaningful') assert exactly this claim. No quantitative cluster-quality metric (e.g., silhouette score, Davies-Bouldin index, gap statistic), no comparison to baseline embeddings (e.g., PCA, t-SNE, or a simple distance on the histograms without the autoencoder), and no permutation/randomization test are provided. With only n=71 and n=33 data points and UMAP n_neighbors=5, apparent clusters can easily arise from sampling noise. I request a quantitative evaluation that separates motion-specific structure from noise, for example by comparing cluster separation of real histograms against histograms with randomly permuted directions, and by reporting a cluster validity index for several hyperparameter choices.","section":"Section 3, Fig. 2, and Section 4"},{"comment":"The pipeline does not isolate breathing motion from correlated nuisance structure. The spherical coordinate frame is defined by the PCA-based lung basis of each patient's lung mask, so the representation is patient-specific in anatomy; the 4D-Lung repeat scans share not only the patient but also scanner, reconstruction protocol, and mask geometry. Therefore, same-patient clustering or visually similar clusters could be driven by these confounds even if the histograms carried zero breathing information. To support the motion-specificity claim, the authors should include a control analysis, for instance an ablation in which the histogram is computed from a randomized or patient-averaged vector field, or an explicit comparison of clustering behavior across different reconstruction binning (phase- vs. amplitude-based) with a statistical test of separation, rather than the current 'was not observed' statement in Section 3.","section":"Section 2.3, Eqs. (1)-(2), and Section 3"},{"comment":"The method contains several free parameters whose influence on the clustering is not investigated: the bin width π/8 for θ and φ, the radius weighting and lung-voxel normalization, the choice of the reference phase j=3, the autoencoder architecture and training hyperparameters, and UMAP's n_neighbors=5, cosine metric, and DensMAP. The Discussion explicitly delegates the angular-resolution question to future research, but the central claim of meaningful clustering is conditional on these choices. At minimum, a sensitivity analysis over a few reasonable bin sizes (e.g., π/4, π/8, π/12) and reference phases, with a cluster-quality metric, is needed to demonstrate that the reported clusters are not an artifact of one arbitrary setting. In addition, the autoencoder is described only as 'lightweight'; its architecture, number of parameters, training epochs, and reconstruction error are not given, which harms reproducibility.","section":"Section 2.3-2.4 and Section 3"}],"minor_comments":[{"comment":"The text says '71 lung cancer patients' for the in-house data set and later '33 4D CT data sets' for 4D-Lung; please clarify whether the in-house number refers to patients or to 4D CT data sets, since the 4D-Lung subset comprises 33 data sets from only 8 patients and the two cohorts are not directly comparable in that respect.","section":"Section 2.1"},{"comment":"The spherical-coordinate equations use x, y, z without defining that these are the components of Vlung; please make the notation explicit (e.g., v = (x, y, z)^T ∈ Vlung) to avoid confusion.","section":"Section 2.3"},{"comment":"The caption and text do not specify the color coding for the in-house UMAP scatter plot; if the top panel is also colored by patient, state this, otherwise describe the markers used. This matters because the reader cannot otherwise distinguish whether the in-house clusters are patient-related or motion-related.","section":"Section 3 and Fig. 2"},{"comment":"The claim that clustering is independent of phase- vs. amplitude-based reconstruction is stated as 'was not observed' without any quantitative measure. If this claim is retained, it should be supported by the same cluster-quality analysis requested in the major comments.","section":"Section 3"},{"comment":"The final paragraph of the Discussion introduces a claimed prior application of the embedding concept to DIR-hyperparameter selection during the Learn2Reg challenges, but gives no reference to the authors' own prior work or challenge report. Please provide a citation or clarify the status of that claim, since it is material to the novelty statement that 'the present publication is the first to explain the underlying methodology.'","section":"Section 4, Learn2Reg paragraph"},{"comment":"The abstract says 'voxel-wise spherical coordinate transformation'; since the transformation is applied to each displacement vector at each voxel (not to the voxel coordinate itself), consider rewording to 'per-vector spherical coordinate transformation' to avoid ambiguity.","section":"Abstract and Section 2.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable feasibility study, but the main claim goes beyond what the evidence supports. The required revisions—quantitative cluster evaluation, confound controls, and hyperparameter sensitivity—are within the scope of a revision and do not require new data collection. I would also flag the Learn2Reg paragraph: it suggests the authors have used this embedding concept in previous challenge work, but no prior publication is cited; the editor may want to check for potential dual submission or undisclosed prior art. Code is promised 'upon acceptance'; ideally it should be made available during revision to allow verification of the autoencoder and UMAP settings."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this if you're thinking about patient-similarity workflows in lung RT. The paper delivers exactly what it promises: a low-cost way to compress DIR vector fields into oriented histograms and embed them with AE+UMAP, and it shows the result on 71 in-house and 33 external 4D CTs. The novelty is real but modest — taking HOF-style histograms into 3D and applying them to registration fields for patient clustering is a legitimate new application, not a new method. The writing is clear and the pipeline is described well enough to reproduce once code is released.\n\nWhat the paper does well: the encoding is simple, the unsupervised pipeline is appropriate for the question, and the authors are honest that this is a feasibility demonstration. The lung-basis normalization for orientation is a sensible step, and the two post-pneumonectomy examples suggest robustness to anatomical outliers, though that's just two cases.\n\nThe soft spot is the evaluation. The central claim that clusters reflect breathing similarity rests on visual inspection of Figure 2 and on repeat scans of the same patient grouping together. Neither observation isolates breathing motion from confounds. The lung-basis is derived from the lung mask shape, so the embedding can capture anatomy. The repeat scans share scanner, protocol, and mask geometry. UMAP with n_neighbors=5 on 71 or 33 points can produce spurious clusters. There is no quantitative cluster metric, no baseline comparison, no permutation test, no stability analysis. The paper mentions the small dataset as a limitation but doesn't address the confound problem.\n\nThat said, the claims in the paper are modest — 'appears to be meaningful' — and the method is a reasonable starting point. I'd send this to review, but the referee should ask for a validation section: silhouette or adjusted Rand index on the repeat-scan labels, a comparison against PCA or random feature embeddings, and a discussion of what the histogram loses by discarding localization. The author's own registration algorithm is from their prior work; that's fine.\n\nWho is this for? Researchers working on motion-based patient stratification, and anyone else who wants a compact vector-field descriptor. It's not a breakthrough, but it's an honest contribution. I'd give it a conditional accept with major revisions.\n\nRecommendation: engage with it, but don't take the clustering claim at face value.","headline":"A tidy, honestly-scoped feasibility study: the new encoding is sensible, but the clustering claim is only supported by anecdote and repeat-scan grouping, not by quantitative validation.","tokens_in":6920,"tokens_out":3099,"would_cite":false,"duration_ms":28245,"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":"The paper claims that breathing motion captured in 4D CT scans can be compressed into compact oriented histograms that cluster patients by breathing-pattern similarity, without any supervised labels.","keywords":["4D CT","deformable image registration","oriented histogram","UMAP","lung radiotherapy","vector field embedding","unsupervised clustering","breathing motion"],"falsifier":"Use the 33 external 4D CTs to compute embedding distances between repeat scans of the same patient and scans of different patients: if same-patient pairs are not systematically closer than different-patient pairs, the meaningful-cluster claim collapses. A second check is to halve the histogram bin size (from $\\pi/8$ to $\\pi/16$) and see whether cluster separation improves; if it does not, the $8\\times16$ representation has already saturated the available motion information.","tokens_in":5929,"feed_emoji":"🫁","tokens_out":8281,"duration_ms":70885,"temperature":0.7,"pith_summary":"The paper tries to establish that high-dimensional deformation vector fields from deformable image registration of 4D CT scans can be reduced to compact oriented histograms, one per motion phase, without losing the patient-specific breathing signature. It reports that after a further autoencoder compression and a UMAP embedding into two dimensions, patients cluster by breathing-pattern similarity rather than by lung size, reconstruction binning, or even the absence of one lung. If this holds, a clinician could compare a new patient's pre-treatment motion data against a database of previous patients, flag unusual motion, or link similar patients to treatment outcomes. The workflow is unsupervised throughout and is demonstrated on 71 in-house and 33 external 4D CT data sets; repeat scans of the same patient mostly landing in the same cluster is the central internal validation.","feed_headline":"Breathing motion maps to 2D clusters for comparing lung patients","feed_subtitle":"Oriented histograms shrink 4D CT motion fields enough to group patients by similar breathing, with no labeled data needed.","key_machinery":"The load-bearing object is the oriented histogram, an extension of histograms of oriented optical flow (a discretized tally of motion directions originally used for action recognition) to 3D displacement fields. The spherical transform $r,\\theta,\\phi$ followed by $r$-weighted binning into an $8\\times16$ array per phase is what converts a vector field of size $[I_x,I_y,I_z,3]$ into a patient-level summary; the autoencoder and UMAP then operate on these summaries. The machinery deliberately discards spatial localization while preserving the distribution of motion directions and magnitudes, and it is this distribution that the cluster analysis links to breathing-pattern similarity.","core_discovery":"The central claim is that a weighted 2D oriented histogram built from the spherical coordinates of each voxel displacement can serve as a compact summary statistic of breathing-motion similarity in lung 4D CT data. Each displacement vector $v=(x,y,z)$ is converted to radius $r$, polar angle $\\theta=\\arccos(z/r)$, and azimuthal angle $\\phi=\\operatorname{arctan2}(y/x)$; vectors inside the lung mask are tallied into bins of size $\\pi/8$ over $\\theta\\in[0,\\pi]$ and $\\phi\\in(-\\pi,\\pi]$, each bin is weighted by $r$, and the result is normalized by lung voxel count. Repeating this for the nine phase-to-reference registrations yields a $9\\times8\\times16$ code per patient. A lightweight unsupervised autoencoder densifies the code, UMAP embeds it in 2D, and the resulting clusters show similar breathing patterns, with most repeat 4D CTs of the same patient clustering together and two post-pneumonectomy patients still falling into interpretable places.","pith_inferences":["I infer the same histogram representation could serve as a retrieval index for treatment-plan libraries, ranking past patients by motion similarity before any dose simulation is run; the paper names this as motivation but does not test retrieval performance.","A natural extension the authors do not report is correlating cluster membership with dosimetric or clinical endpoints; if clusters track motion similarity, they may also track motion-induced target-miss risk, but this connection is my inference, not their result.","Because localization is discarded, I expect the method to be blind to spatial rearrangements of motion (for instance, a whole-tumor translation versus a local diaphragm spike), so a multiscale version that bins displacement fields in subregions is a direct testable upgrade.","The visual cluster analysis suggests an obvious quantitative check, such as repeat-scan nearest-neighbor accuracy or silhouette score, which the paper leaves for future work; I infer such a score would make the method easier to trust and tune."],"forward_implications":["A new patient's 4D CT can be embedded into the same UMAP space using only pre-treatment planning data, enabling prospective comparison with previously treated patients without any supervised training.","The histogram encoding is not lung-specific, so the same reduction and embedding pipeline applies to other registration tasks, anatomies, or physiological processes.","Because repeat 4D CTs of the same patient mostly cluster, the embedding is stable enough to serve as a longitudinal check on whether a patient's breathing motion is changing between scans.","Any deformable registration algorithm's output can be fed into the pipeline, and the paper reports using the embedding to select patient-specific registration hyperparameters based on nearest neighbors in the UMAP space.","Clustering appears independent of whether the 4D CT was reconstructed by phase-based or amplitude-based binning, so the representation captures motion characteristics beyond the reconstruction protocol."],"supporting_citations":[{"why":"Supplies the variational deformable registration method and hyperparameters that produce the vector fields encoded by the histograms.","marker":"[6]"},{"why":"Provides the histogram-of-oriented-optical-flow idea that the paper extends to 3D vector fields.","marker":"[7]"},{"why":"Justifies the autoencoder-before-manifold-learning design by showing it improves cluster quality.","marker":"[8]"},{"why":"Defines the UMAP algorithm that produces the 2D embeddings and clusters.","marker":"[9]"},{"why":"Adds density-preserving DensMAP regularization used in the UMAP embedding.","marker":"[10]"},{"why":"Supplies the external 4D Lung data set with repeat 4D CTs, which provides the independent test data and the same-patient clustering evidence.","marker":"[3–5]"}],"fun_headline_variants":["4D CT motion shrunk to 2D clusters for lung patient matching","Oriented histograms turn lung motion into patient clusters","Breathing maps to 2D clusters via oriented histograms","Vector fields to 2D: clustering lung patients by motion","Compact motion signatures cluster lung 4D CT patients"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a fixed $8\\times16$ histogram of displacement directions, weighted by vector length and normalized by lung voxel count, keeps enough patient-specific motion information that distances in the UMAP embedding reflect true breathing similarity; this is checked only by visual cluster inspection, with no quantitative separation measure.","fun_headline_variants_meta":{"raw":{"variants":["4D CT motion shrunk to 2D clusters for lung patient matching","Oriented histograms turn lung motion into patient clusters","Breathing maps to 2D clusters via oriented histograms","Vector fields to 2D: clustering lung patients by motion","Compact motion signatures cluster lung 4D CT patients"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000162,"raw_usage":{"total_tokens":1276,"prompt_tokens":1020,"completion_tokens":256,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":636,"completion_tokens_details":{"reasoning_tokens":171}},"tokens_in":636,"tokens_out":256,"duration_ms":2865,"temperature":1.0,"reasoning_tokens":171,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:16:13.222905+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use the 33 external 4D CTs to compute embedding distances between repeat scans of the same patient and scans of different patients: if same-patient pairs are not systematically closer than different-patient pairs, the meaningful-cluster claim collapses. A second check is to halve the histogram bin size (from $\\pi/8$ to $\\pi/16$) and see whether cluster separation improves; if it does not, the $8\\times16$ representation has already saturated the available motion information.","supporting_citations":[{"cited_title":"The data evaluation was approved by the local ethics board, and the requirement to obtain written informed consent was waived [WF-82/18] between 2015 and 2020","cited_arxiv_id":null,"evidence_quote":"Supplies the variational deformable registration method and hyperparameters that produce the vector fields encoded by the histograms."},{"cited_title":"Technological quality requirements for stereotac- tic radiotherapy,","cited_arxiv_id":null,"evidence_quote":"Provides the histogram-of-oriented-optical-flow idea that the paper extends to 3D vector fields."},{"cited_title":"Cor- respondence model-based 4d vmat dose simulation for analysis of local metastasis recurrence after extracranial sbrt,","cited_arxiv_id":null,"evidence_quote":"Justifies the autoencoder-before-manifold-learning design by showing it improves cluster quality."},{"cited_title":"Evaluation of 4-dimensional computed tomography to 4-dimensional cone-beam computed to- mography deformable image registration for lung can- cer adaptive radiation therapy,","cited_arxiv_id":null,"evidence_quote":"Defines the UMAP algorithm that produces the 2D embeddings and clusters."},{"cited_title":"Data from 4D lung imaging of NSCLC patients,","cited_arxiv_id":null,"evidence_quote":"Adds density-preserving DensMAP regularization used in the UMAP embedding."}],"review_version":1}