{"id":"3654cb40-52d5-4c95-93aa-d4234f4f29c9","arxiv_id":"2608.00145","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A single CMR slice can localize the left ventricle and regress its continuous 3D short-axis orientation, with a proposed PCC embedding achieving 3.39° average angle error across four datasets.","lead":"This paper proposes a deep-learning method that, from one arbitrary cardiac MRI slice, finds the left ventricle and estimates the 3D orientation of the heart's short-axis plane. It also introduces a new orientation-encoding scheme and a synthetic benchmark built by slicing 3D heart scans.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Benchmark samples only planes through the LV mask, so the 'arbitrary CMR slice' claim is unsupported; the 4-volunteer real-world validation is too small to close this gap.","rationale":"The reader's weakest assumption — that the synthetic slicing benchmark is representative of arbitrary clinical CMR slices — is precisely the most load-bearing condition for the paper's central claim. The pipeline guarantees LV visibility by construction, so the benchmark only measures performance on slices that contain the target. The title and abstract promise 'arbitrary CMR slice', which in clinical practice includes localizers, 2CH/4CH approximations, and slices with no LV. The 4-volunteer real-world test is too small to establish generalization, especially since it may not include negative cases. Without a broader real-slice evaluation, the reported accuracy is an upper bound for a favorable subsample. Secondary issues raised by the reader — test-set tuning on V2S-Real and the Eq. (1) unit-norm typo — are real and should be corrected, but they do not change the overall conditional verdict. The central method is plausible, the adapted baselines are reasonable, and the experiments are internally consistent on the constructed benchmark; the weakness is external validity. Thus the reader's CONDITIONAL verdict remains appropriate.","tokens_in":8367,"tokens_out":5792,"duration_ms":66390,"concrete_test":"Take an independent set of real CMR volumes with full-stack or localizer slices and DICOM metadata (e.g., ACDC test volumes or M&Ms-2 test data), render slices exactly as acquired — including non-LV and partial-LV slices — run the released model without retraining, and measure: (a) false-positive detection rate on slices with no LV, (b) angle deviation on slices where an LV is present, and (c) localization mIoU restricted to detected boxes. If the false-positive rate is non-negligible or angle error degrades substantially relative to Table 1, the 'arbitrary CMR slice' claim is not supported by the current evaluation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a single arbitrary CMR slice suffices for LV localization and 3D SAX orientation estimation. The benchmark (Sec. 3.4) generates every slice by first sampling a point P inside the LV mask, then sampling a plane through P. This guarantees the LV is visible and roughly centered in every training and test slice; slices without LV, with partial LV, or with clinically typical localizer geometry are excluded. The model is therefore evaluated only on the conditional distribution P(slice | LV visible), not on arbitrary clinical slices. The real-world comparison (Fig. 5) uses only 4 volunteers and appears to include no explicit negative cases without LV. Consequently, the reported Table 1 numbers — 86.18% mIoU and 3.39° average angle deviation — cannot be read as evidence for the general 'arbitrary slice' claim. The paper's task framing requires that the model handle the full clinical range of slice positions and orientations, but the benchmark's slice-sampling distribution is narrower than that.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a new task: given a single 2D CMR slice, jointly localize the left ventricle (as a rotated 2D box) and estimate the 3D orientation of the short-axis plane. It adapts four orientation-aware detection methods (RIoU3D, FCOS3D, GWD3D, PSC) to this setting, analyzes their limitations, and proposes Polar-Coupled Circular (PCC) embedding: a circular-harmonic representation of the SAX normal coupled with the polar component. A benchmark is constructed by randomly slicing 3D CMR segmentation volumes through the LV mask, and experiments on ACDC, M&Ms-2, In-House, and V2S-Real report average mIoU 86.18%, average angle deviation 3.39°, and AP75 87.96%. A real-world SAX plane prescription experiment on four volunteers is also reported.","tokens_in":8568,"tokens_out":5675,"duration_ms":67742,"significance":"If the results hold, the paper opens a useful direction for automated cardiac plane prescription and cross-view analysis: estimating a continuous 3D orientation from a single slice is more general than discrete view classification and does not require multi-slice geometric intersection. The PCC embedding is geometrically motivated, and the paper includes reproducible assets: code is released, the benchmark generation pipeline is described in detail, and orientation labels are derived from external known SAX geometry, so there is no circularity in the central evaluation. The comparison with four adapted baselines is also informative. The main risks are the mismatch between the benchmark slice-sampling distribution and truly arbitrary clinical slices, the very small real-world validation, and a typographical issue in Eq. (1) that, as printed, breaks the central unit-energy property.","major_comments":[{"comment":"As typeset, Eq. (1) defines p_n = sqrt(2/(N-1)) (1 - n_z^2) cos(θ + 2πn/(N-1)). With this definition, ||u||_2^2 = (1-n_z^2)^2 + n_z^2, which is not identically 1, contradicting the claimed unit-energy constraint. The reconstruction in Eq. (3) then recovers (1-n_z^2)(cos θ, sin θ) rather than (n_x, n_y). The missing factor appears to be sqrt(1-n_z^2); with that factor, the sum of squares of the circular components equals 1-n_z^2 and the full embedding has unit norm, and Eq. (3) correctly recovers the normal. Please correct Eq. (1) and verify the reconstruction, or explicitly justify a different intended definition.","section":"§3.1–3.2, Eq. (1)–(3)"},{"comment":"The benchmark is built by first sampling a point P inside the LV mask and then sampling a plane through P. This guarantees that every training and test slice intersects the LV, so the model is evaluated only on the conditional distribution P(slice | LV is visible). The title, abstract, and introduction claim 'arbitrary CMR slice,' but non-LV slices, severely partial LV slices, and clinically typical localizer views are excluded. The real-world validation in Fig. 5 uses only four volunteers and does not include negative or non-LV slices. Therefore Table 1's numbers cannot be read as evidence for the general 'arbitrary slice' claim. I ask for either (i) additional evaluation on a retrospectively collected set of true clinical slices, including slices that do not intersect the LV, or (ii) a clearly qualified statement that the task is single-slice SAX-plane estimation conditional on LV pres","section":"§3.4 and Fig. 5"},{"comment":"The In-House row for FCOS3D is identical to the M&Ms-2 row (4.94±3.84, 84.22, 99.78, 81.44). This appears to be a copy/paste error. Since Table 1 is the central quantitative evidence for the cross-dataset improvement, this duplicated row must be corrected or recomputed, and any averaged performance numbers that depend on it must be updated.","section":"Table 1, In-House row"},{"comment":"The multi-view comparison is based on only four volunteers, and the figure reports only aggregate mean±std. No per-volunteer measurements, number of localizer slices, acquisition details, or statistical tests are provided. With n=4, the claim of being 'on par with' or better than multi-view methods is not established. Please report the underlying individual measurements and a statistical comparison, or present this as a preliminary pilot result.","section":"§4, Fig. 5c"}],"minor_comments":[{"comment":"The decoding step does not state whether the recovered normal is renormalized after projection. Once Eq. (1) is fixed this may be unnecessary, but the paper should state explicitly whether any post-processing normalization is applied.","section":"§3.2"},{"comment":"For RIoU3D and GWD3D, the 180° rotational ambiguity means the network can predict either n or -n. Reporting angle deviation with this sign ambiguity conflates 'reversed prediction' with geometric error; consider also reporting the minimal axis-angle deviation for these baselines to clarify the failure mode.","section":"§2 / Table 1"},{"comment":"The text says '20 slices are sampled from each mask-annotated 3D CMR volume,' but then V2S-Real is described as containing 2,000 frames. It would be helpful to state the resulting total number of training/test slices per dataset and whether slices from the same volume appear in both training and test sets.","section":"§3.4"},{"comment":"The In-House subset description says 188 scans with 95 for testing, but does not explain how the training split is formed or whether there is patient overlap between training and test. Please clarify this for reproducibility.","section":"§4"},{"comment":"The term 'arbitrary CMR slice' is used much more broadly than the actual experimental setting. I recommend consistently qualifying the claim as 'arbitrary slice intersecting the LV' unless the broader evaluation is added.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising idea and a well-structured study, but the main claim about arbitrary slices is not yet supported by the benchmark design. The Eq. (1) issue is likely a simple typo, but it is central to the proposed embedding and must be fixed. The duplicated FCOS3D row in Table 1 also needs correction before the manuscript can be considered acceptable. I would be willing to review a revised version that addresses these points."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The task here is genuinely new: regress the continuous 3D normal of the LV short-axis plane from a single 2D CMR slice, while localizing the LV with a rotated box. Prior work stops at discrete view classification or needs multiple slices. The PCC embedding is a clean extension of circular-harmonic encodings to 3D with a unit-energy constraint, and their adaptation of four detection baselines (FCOS3D, RIUo3D, GWD3D, PSC) gives a useful picture of where those fail. The benchmark pipeline that resamples planes from existing segmentation volumes is practical and will be reused.\n\nThat said, the central claim is oversold. Every training and test slice is generated by sampling a plane through a randomly chosen point inside the LV mask, so the ventricle is always visible and roughly centered. The evaluation therefore concerns the conditional distribution 'slice contains LV', not 'arbitrary CMR slice'. The four-volunteer real-world prescription test is encouraging but too small to close the gap. The paper also selects N=4 and w_emb=2 by ablation on V2S-Real, which then appears in the main test table — textbook test-set tuning, though the ablation differences are small enough that the conclusion likely survives.\n\nTwo smaller points: Table 1 shows the same FCOS3D numbers for M&Ms-2 and In-House, which is almost certainly a copy-paste error. And Eq. (1), if read with the square root over the full factor, does yield unit norm (the sum of (N-1) equally spaced cos^2 terms is (N-1)/2), so I would not flag that as an error.\n\nOverall, the core idea is sound, the method is sensible, and the benchmark is a real contribution. The paper deserves a serious referee. My main request would be to either soften the 'arbitrary' language or add an evaluation on slices that miss the LV, and to fix the minor editorial issues.","headline":"New single-slice CMR orientation task with a sound method, but the 'arbitrary slice' claim outruns the benchmark's LV-centered slice distribution.","tokens_in":9083,"tokens_out":4001,"would_cite":true,"duration_ms":39851,"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 arbitrary 2D CMR slice is enough to localize the left ventricle and recover the full 3D orientation of the short-axis plane, using a continuous unit-energy embedding.","keywords":["left ventricle localization","CMR orientation estimation","short-axis plane","polar-coupled circular embedding","oriented object detection","single-slice regression","cardiac MRI","rotated bounding box"],"falsifier":"Benchmark the trained model on a new cohort of, say, 50 patients whose SAX normals are taken from DICOM metadata of clinically acquired single-slice localizers. If the average angle deviation on those real slices rises substantially above the reported 3.39°, the single-slice recoverability claim would be refuted. A second probe: feed the model slices that do not intersect the LV mask; if it confidently predicts an orientation rather than reporting absence, the localization-orientation coupling is incomplete.","tokens_in":8221,"feed_emoji":"🫀","tokens_out":3474,"duration_ms":35753,"temperature":0.7,"pith_summary":"The paper is trying to establish a new task: from one arbitrary 2D cardiac MRI slice, a network can both locate the left ventricle and estimate the 3D direction of the short-axis plane. It argues that existing orientation representations — view classification, direct angle regression, IoU/Gaussian overlap, and the phase-shifting coder — fail on this task due to 180° ambiguity, boundary discontinuity, or instability near the poles. The proposed Polar-Coupled Circular (PCC) embedding encodes the plane normal as a unit-norm circular-harmonic vector that is continuous, unambiguous, and geometrically consistent. If right, this would allow SAX plane prescription and spatially aware analysis from a single slice rather than multi-view intersection. Reported results are an average mIoU of 86.18% and an average angle deviation of 3.39° across four datasets.","feed_headline":"One MRI slice reveals the heart's 3D short-axis plane","feed_subtitle":"A unit-energy embedding turns single-slice LV orientation into a stable regression, hitting 3.39° average error.","key_machinery":"The Polar-Coupled Circular (PCC) embedding: for a unit normal n, the embedding coefficients are p_n = sqrt(2/(N-1)) * sqrt(1-n_z^2) * cos(theta + 2πn/(N-1)) plus p_{N-1}=n_z. The key property is that all coefficients together have unit Euclidean norm, and the azimuthal components are scaled by sqrt(1-n_z^2), so as the plane normal approaches the poles the azimuthal information smoothly fades out. This gives a continuous, one-to-one, boundary-free parameterization of the semi-sphere, which lets a standard FCOS-style detector regress orientation with ordinary L1 loss.","core_discovery":"The paper's central claim is that 3D SAX orientation can be regressed directly from a single 2D CMR slice, and that the PCC embedding makes this regression stable in ways that previous rotation representations are not. The ground-truth SAX unit normal is encoded by the PCC map into N circular-harmonic coefficients with unit-energy norm; the network regresses these coefficients, and a linear projection decodes them back to a Cartesian normal. The authors show this beats four adapted orientation-aware detectors in both orientation accuracy and localization consistency, with lower standard deviation, across four datasets spanning different centers, vendors, and cardiac phases.","pith_inferences":["The same unit-energy circular-harmonic encoding applies to any task where a 3D plane normal must be regressed from a 2D view — e.g., other standard cardiac planes (2CH/4CH) or fetal/neonatal imaging — so the mechanism is not LV-specific.","If truly arbitrary clinical slices include planes that miss the ventricle, the detector would need an explicit 'no LV' output; the current benchmark guarantees LV visibility by construction, so rejection behavior remains untested.","The small real-world validation (four volunteers) leaves open how the performance transfers to patients with pathology or unusual anatomy; a larger prospective study would close that gap."],"forward_implications":["SAX plane prescription could be performed from a single localizer slice, removing the need for multi-slice geometric intersection.","Orientation predictions over the full semi-sphere are unambiguous: 180° flips that plague IoU/Gaussian-based methods are avoided.","The synthetic slicing pipeline turns commonly available 3D segmentation datasets into large-scale orientation training data without manual annotation.","The method generalizes across multi-center and multi-vendor datasets and across cardiac phases (In-House 30-frame annotation), suggesting phase-invariant orientation cues."],"fun_headline_variants":["One CMR slice yields LV short-axis orientation with 3.4° error","Single-slice CMR embedding nails 3D LV orientation at 3.39°","PCC embedding estimates LV short-axis plane from one MRI slice","Single CMR slice predicts LV 3D orientation with 3.39° accuracy","Regression from one MRI slice gives LV short-axis orientation"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The assumption that a single 2D CMR slice containing part of the left ventricle carries enough information to determine the full 3D orientation of the short-axis plane, with the random planes sampled through LV masks in segmentation volumes treated as representative of arbitrary clinical slices.","fun_headline_variants_meta":{"raw":{"variants":["One CMR slice yields LV short-axis orientation with 3.4° error","Single-slice CMR embedding nails 3D LV orientation at 3.39°","PCC embedding estimates LV short-axis plane from one MRI slice","Single CMR slice predicts LV 3D orientation with 3.39° accuracy","Regression from one MRI slice gives LV short-axis orientation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000714,"raw_usage":{"total_tokens":3039,"prompt_tokens":728,"completion_tokens":2311,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":472,"completion_tokens_details":{"reasoning_tokens":2212}},"tokens_in":472,"tokens_out":2311,"duration_ms":16539,"temperature":1.0,"reasoning_tokens":2212,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T01:09:42.362216+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Benchmark the trained model on a new cohort of, say, 50 patients whose SAX normals are taken from DICOM metadata of clinically acquired single-slice localizers. If the average angle deviation on those real slices rises substantially above the reported 3.39°, the single-slice recoverability claim would be refuted. A second probe: feed the model slices that do not intersect the LV mask; if it confidently predicts an orientation rather than reporting absence, the localization-orientation coupling is incomplete.","supporting_citations":[],"review_version":1}