acceptodds
Under review as a conference paper at ICLR 2027

When do coordinates help projection-to-volume learning?

Abstract

Positional coordinates can resolve spatial ambiguity in learned 3D inverse problems, but their benefit can reverse as measurement coverage increases. We study this in 3D lung segmentation from simulated X-ray projections, separating position from calibrated lifting, attention, and decoder processing. On 24 held-out LCTSC patients under a prospectively locked protocol, coordinates increase source-acquisition Dice by 30.8 percentage points at one view but reduce it by 6.3 at twelve across three initializations. Two-view source acquisitions show a 19.5-point benefit when clustered and a 3.6-point penalty when spread. In our pipeline, sparse-view gains persist without learned attention or the separate geometry-conditioning branch, but accuracy drops substantially without both coordinates and learned spatial slots. Across two development cohorts, along-ray depth exceeds detector-plane coordinates by 9.7 and 3.1 source-acquisition Dice points across three initializations, with advantages persisting under held-out rotations, consistent with difficulty localizing along projection rays. Synthetic coordinate benefits diminish or disappear for uniformly translated populations and depend on sampled frequencies. In a matched context experiment, wider decoder context recovers nearly half of the single-view coordinate advantage without positional input. A scalar acquisition-dependent gate preserves prespecified sparse and broad endpoints in both development cohorts; broader training coverage only partly repairs its intermediate-angle deficit. Coordinate utility therefore depends jointly on measurement ambiguity, population alignment, and decoder processing; neither view count nor geometry alone predicts its benefit.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
positional information, inverse problems, inductive biases, geometric deep learning, sparse-view imaging, anatomical segmentation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.80 over 5 reviews, above 27% of the others.
Reviewer_duBNOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies when coordinate input is helpful for learned projection-to-volume models. The main idea is that coordinates work like a population prior, so they help a lot when the measurements are ambiguous (for example single view, where the depth along the ray is unknown), but they become not useful or even harmful when the measurements are enough. The authors separate the effect of position from other spatial components like calibrated lifting, attention and decoder context, and also check which direction of position is more useful and in which conditions it helps. They validate this mainly on 3D lung segmentation from simulated X-ray projections, including one held-out cohort with locked protocol, and on some synthetic inverse problems. They propose a simple gate that changes the coordinate strength by acquisition geometry, which works for very sparse and broad views but not for intermediate ones.

Critical strengths
  • The main OFF/FULL reversal is supported better than usual for this kind of empirical study. The LCTSC protocol is fixed before training, the labels are manual contours, the scan box is not cropped by label, and the sign holds in all three seeds with paired-patient intervals clearly away from zero. This is important for my assessment, because it shows the core observation is not only an artifact of the reused development cohorts or the automated labels.
Critical weaknesses
  • It is not clear for me that the reversal is a property of positional input under broad coverage, rather than a property of training one model on mixed coverage. All clinical OFF/FULL comparisons train a single model on 1 to 12 views together, and for the sparse samples the coordinate is the main localization cue. The scalar gate result actually points to this direction. With the same mixed training, removing position only for broad inputs recovers OFF-level broad accuracy, and FULL also fits the twelve-view training cases worse than OFF (0.838 vs 0.859, Table 15). Without OFF/FULL models trained on broad coverage only, the claim that coordinate benefit reverses as coverage increases is not yet separated from this training setup.

  • The paper has no prior-only reference, for example a coordinate-only model without image input or simply the training mean mask. I think this matters a lot, because on LCTSC the FULL model changes very little with more views (0.672 at one view to 0.740 at twelve, while OFF goes from 0.364 to 0.803). This suggests FULL mostly outputs a population shape and makes little use of additional measurements. If this is the case, both the sparse gain and the broad penalty can be explained by a simple prior-versus-evidence tradeoff, and the depth, synthetic and context analyses add limited understanding beyond it. The proposed gate is also closer to a negative result. With the cutoff 0.25, the formula in Appendix D gives a zero gate exactly at a 60 degree pair, so the intermediate-span failure is expected by construction rather than a new finding.

Questions for authors
  • How is the patient box, and therefore c0 and the detector scale, defined for LIDC-IDRI and TotalSegmentator? Appendix H says LCTSC uses the full scan box without a label-defined crop, so I want to check whether the development cohorts use a label-derived box. This affects how I read the development-only results (gate, depth vs plane, context).

  • Section 5.1 says the full and no-coordinate arms were not repeated in the directional replication. Were they run in the original single-seed directional experiment? If so, please report them, because without OFF it is hard to tell whether depth helps or detector-plane coordinates hurt.

  • Could you train OFF and FULL on broad coverage only (for example four spread or twelve views) and report whether the broad penalty still remains?

  • Could you report one-view and twelve-view Dice of a coordinate-only model without image input, or of the training mean mask, at least on LIDC-IDRI and LCTSC?

  • The related work misses sparse-view CBCT methods that combine calibrated back-projection queries with coordinate encodings, such as DIF-Net (MICCAI 2023) and C2RV (CVPR 2024). Discussing them would help to position the study.

5 comments
AuthorsOpenReview ↗

Response to Reviewer duBN

Q: It is not clear that the reversal is a property of positional input under broad coverage, rather than a property of training one model on mixed coverage. All clinical OFF/FULL comparisons train a single model on 1 to 12 views together. Could you train OFF and FULL on broad coverage only (for example four spread or twelve views) and report whether the broad penalty still remains?

A: We ran this experiment, and it supports your reading. We trained OFF and FULL on twelve views only, on both cohorts, with the same optimizer settings and update budget as the mixed-view models. FULL−OFF at twelve views, 11,250 updates, three-seed means on the evaluation patients, with patient-bootstrap 95% intervals:

  • LCTSC, mixed-view training: −11.28 Dice points [−14.44, −8.35]; the difference on the training patients is +0.51.

  • LCTSC, twelve-view-only training: −0.10 [−0.44, +0.22].

  • LIDC-IDRI, mixed-view training: −1.43 [−2.26, −0.83]; the difference on the training patients is −0.10.

  • LIDC-IDRI, twelve-view-only training: −0.02 [−0.07, +0.03].

The dense-view penalty is therefore largely tied to mixed-view training. We have changed the claim accordingly: a model trained across view counts with a coordinate input loses dense-view accuracy; we do not claim that coordinates are harmful under broad coverage as such.

One limitation remains. At equal update counts, twelve-view-only training sees five times as many twelve-view examples and 2.22 times as many projections, so this control changes exposure as well as the training distribution. It does not isolate mixed-view interference. Section 5.1, the abstract and the conclusion of the revised paper now report this control and state the narrower claim. We therefore also matched twelve-view exposure instead of updates, by comparing mixed-view models at 11,250 updates with twelve-view-only models at 2,250. The regime difference remains on LCTSC (+11.45 points [8.46, 14.71]) but not on LIDC-IDRI (−0.57 [−1.30, 0.25]). Exposure alone therefore does not explain the LCTSC contrast, and the cause remains unresolved. Table 2 in Section 5.1 and Appendix I.1 give these results. A single-seed DIF-Net adaptation trained on mixed views shows no twelve-view penalty (Table 5), so the effect is specific to our pipeline.

AuthorsOpenReview ↗

Response to Reviewer duBN

Q: The paper has no prior-only reference, for example a coordinate-only model without image input or simply the training mean mask. On LCTSC the FULL model changes very little with more views (0.672 at one view to 0.740 at twelve, while OFF goes from 0.364 to 0.803). This suggests FULL mostly outputs a population shape. Could you report one-view and twelve-view Dice of a coordinate-only model or of the training mean mask?

A: We computed the mean-mask reference and trained a coordinate-only model. The training mean mask scores 55.37% Dice on the LCTSC test patients (95% interval 51.28–59.51%) and 72.59% on the LIDC-IDRI development patients (69.43–75.49%); it is the same at every view count.

  • In the original runs, FULL is above the mean mask at one view (67.19% on LCTSC, 80.84% on LIDC-IDRI) and at twelve views (73.98% and 91.36%). The matched coordinate-only comparison below further supports a benefit from image conditioning at the long budget, although these comparisons do not quantify how much FULL relies on patient-specific projection evidence.

  • One-view OFF is below the mean mask on both cohorts (36.37% and 49.74%). Without coordinates, one view therefore yields less than average anatomy, which is consistent with your reading. The comparison does not establish what fraction of the gain is a shape prior.

  • On LCTSC, FULL gains 6.8 points from one to twelve views while OFF gains 44.0, which fits your prior-versus-evidence reading. On LIDC-IDRI FULL gains 10.5 points.

Section 5.1 of the revised paper now reports both image-free references. Figure 1 plots this contrast on the LCTSC test patients, and Figure 4 shows the mean mask and a coordinate-only model next to the OFF and FULL predictions for two of them. The coordinate-only model receives no image and scores 57.46% on LCTSC and

75.12% on LIDC-IDRI at every view count. In the matched long-budget runs (Appendix I.2, Table 17), FULL exceeds it at one view (68.16% and 80.37%) and at twelve views (72.53% and 91.42%), while one-view OFF is below it (38.90% and 53.47%). At the short budget, LIDC-IDRI one-view FULL (68.99%) is below both references, so the result depends on the training budget. The coordinate-only model bypasses the image branch entirely. The encoder and lifting operations are skipped; their parameters remain stored but frozen. We do not feed zero-valued projections through the encoder. Patient-centered coordinates pass through the coordinate MLP directly to the 3D decoder, and only these two components are trained. Training follows the same patient order, mixed-view schedule, loss and optimizer as the corresponding local FULL model. Both use three matched seeds, with checkpoints at 2,250 and 11,250 updates. The coordinate-only prediction is independent of view count, but its physical coordinates still provide patient-box size information. Appendix I.2 states this.

Q: The proposed gate is closer to a negative result. With the cutoff 0.25, the formula in Appendix D gives a zero gate exactly at a 60 degree pair, so the intermediate-span failure is expected by construction rather than a new finding.

A: We agree, and the revised paper presents the gate as a test with a negative outcome between the endpoints. The contribution is retitled “A test of geometry-based gating”. Appendix D states that the gate reaches zero at 60° and treats layouts identically even though their measured content differs. Section 5.3 now says that the 60° gate is zero, so the rule removes the coordinate input in a layout where FULL (0.853) outperforms OFF (0.760), and that SCALAR (0.761) empirically scores close to OFF. Appendix D adds that this closeness is an empirical result and not a consequence of the rule alone, because SCALAR and OFF are trained separately. Most of the gate material has moved to Appendix D; the main text keeps the definition (Section 3, Equation 4) and one results paragraph (Section 5.3).

AuthorsOpenReview ↗

Response to Reviewer duBN

Q: How is the patient box, and therefore c0 and the detector scale, defined for LIDC-IDRI and TotalSegmentator? Appendix H says LCTSC uses the full scan box without a label-defined crop, so I want to check whether the development cohorts use a label-derived box.

A: We audited the patient boxes by replaying the stored geometry (Appendix I.7). For LIDC-IDRI (215 cases) and LCTSC (60 cases, from the native CT headers), the patient box is the full CT volume, not a crop derived from the organ masks, and c0 is the centre of that box. The maximum corner discrepancy in the replay is

0.00002 mm for LIDC-IDRI and 0 mm for LCTSC. TotalSegmentator also uses the full CT-volume box, not a box derived from segmentation labels. Its bounds are calculated from the CT dimensions and spatial affine. Labels are used for organ-volume eligibility checks when selecting scans, but not to define the box. The audit of this cohort covers 248 processed records, of which six were replayed from native data. For all three cohorts, we project the eight CT-volume corners onto the detector and choose one pixel spacing large enough to contain them across the specified angles, with a 10% margin. The same spacing is used in both detector directions and for every view of that patient. LIDC-IDRI uses the original 12 angles, spaced 15° apart, and this spacing is retained when the bank is expanded to 24 views. LCTSC and TotalSegmentator use all 24 angles, spaced 7.5° apart. The detector scale therefore depends on the CT volume and imaging geometry, not the segmentation labels. Section 3 now states the box rule, and Appendix I.7 states both the box and the detector-scale rules.

Q: Section 5.1 says the full and no-coordinate arms were not repeated in the directional replication. Were they run in the original single-seed directional experiment? If so, please report them, because without OFF it is hard to tell whether depth helps or detector-plane coordinates hurt.

A: We have now run both arms. Section 5.2 and Appendix I.4 (Table 19) add a matched OFF model and a model with all three acquisition-frame coordinates to the directional comparison on LIDC-IDRI, with three seeds. At source acquisitions, depth improves over OFF by 22.93 points [21.62, 24.14], detector-plane coordinates by 13.25 [11.93, 14.49], and both together by 23.42. Depth therefore helps, and detector-plane coordinates also help, by less; they do not hurt relative to OFF. The anchors have not been repeated on TotalSegmentator. To answer your question directly: yes, the original seed-0 run included both arms. On the 37 LIDC-IDRI development patients after 11,250 updates, using heart-excluded lung Dice, OFF scores 0.5984 (source), 0.5849 (held rotations) and 0.5917 (offset), and the model with full acquisition coordinates scores 0.8320, 0.8297 and 0.8256. Here FULL uses acquisition-frame coordinates, not the world-coordinate input of the main FULL model, and these are original single-seed results, not three-seed averages. Appendix I.4 now reports them.

Q: The related work misses sparse-view CBCT methods that combine calibrated back-projection queries with coordinate encodings, such as DIF-Net (MICCAI 2023) and C2RV (CVPR 2024).

A: Thank you for these references. We have added DIF-Net and C2RV to Section 2 (Related Work) and cite them in the first paragraph of the introduction. Both are close to our pipeline: they project each 3D query point onto the views and regress the output from the image features sampled there. C2RV additionally interpolates multi-scale 3D feature volumes and fuses the views with cross-attention. One clarification from our reading: in both papers the query point is used only for projection and interpolation, and its coordinate is not given to the network as an input feature (DIF-Net, Section 2.2; C2RV, Section 3, Eqs. 1, 2, 7 and 9). In our terms both correspond to the OFF setting, and both are evaluated at 6 to 10 views. We also tested the coordinate input in our own adaptation of DIF-Net (Section 5.4, Table 5). There it improves Dice at every view count, including twelve views, so the broad-view penalty of our pipeline does not appear in that architecture. The cited method that does add a positional encoding of the coordinate, PerX2CT, reports a gain from it at two views, which agrees with our sparse-view result. We will gladly correct this reading if we have missed a coordinate input in either method.

Reviewer_K6kXOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the utility of explicit coordinates in learned reconstruction of three-dimensional lung masks from simulated, calibrated X-ray projections. Within a shared lifting-and-decoding pipeline, it separates explicit position from calibrated routing, learned aggregation, spatial slots, and decoder context. It finds that coordinates can strongly help under sparse or clustered measurements while harming performance under broader coverage. Directional and synthetic experiments examine along-ray ambiguity, population alignment, and spatial processing. A prospectively locked LCTSC experiment with manual contours and 24 held-out patients reproduces coverage-dependent effects across three initializations. A simple geometry-dependent coordinate gate preserves selected endpoints but does not resolve intermediate acquisition layouts. The contribution is an empirical analysis of positional inductive bias, rather than a new state-of-the-art reconstruction architecture.

Critical strengths
  1. This paper asks an important question that is often ignored: positional cues may help at a sparse-view operating point yet become redundant or harmful when measurement coverage changes. The LCTSC sign reversal is strong evidence for this pipeline and population.

  2. The central finding is supported by a meaningful independent replication. On the locked LCTSC cohort, coordinates improve source-acquisition Dice by 30.81 points at one view and 19.55 points at two clustered views, but reduce it by 3.63 points at two spread views and 6.34 points at twelve views. The equal-view-count comparison is particularly informative. Patient-level paired intervals and separate seed variation are reported, with repeated acquisitions correctly distinguished from independent patients.

  3. The study separates several commonly conflated sources of spatial information. Calibration remains present when explicit coordinates or learned aggregation are disabled; directional inputs, translated synthetic populations, and decoder-context changes probe different explanations. The reporting of gate failures and reused development cohorts strengthens the interpretation. Relative to CoordConv, positional-information probes, and geometry-aware reconstruction such as PerX2CT, the useful increment is the systematic study of measurement coverage and its interactions.

Critical weaknesses
  1. The results establish pipeline-specific coordinate utility more strongly than a general mechanism. The clinical studies largely share additive coordinate fusion and a particular coarse lifting/decoder design. FULL also fits twelve-view training cases worse than OFF (Table 15), so the negative test effect cannot be assigned solely to a harmful population prior or generalization. This is acknowledged, but it leaves optimization and fusion interference as substantive alternative contributors. The current evidence does not establish that the same sign reversal would persist with other positional interfaces or converged training.

  2. The explanatory interventions have uneven replication. The independent LCTSC study tests OFF/FULL coverage dependence, whereas the original full context-by-coordinate interaction is single-seed and the directional analyses remain on reused development cohorts. The subsequent multi-seed coordinate-free context comparisons are useful but do not replicate the full interaction. Likewise, the two-spread-view effect does not have the same sign across cohorts. These limitations reduce the strength of a broadly predictive account. Actually, the FULL-OFF contrast is an end-to-end comparison between separately trained systems, not a direct estimate of a coordinate signal’s intrinsic effect. Meanwhile, coordinates alter optimization and effective function class. Hence, the conclusion should remain specific to the tested pipeline and training protocol.

Questions for authors
  1. Given the worse twelve-view training fit of FULL, what do the existing diagnostics distinguish between measurement-dependent positional bias and optimization/interference introduced by the additive coordinate pathway? Please identify which mechanistic conclusion is supported beyond the demonstrated fixed-budget, pipeline-specific effect.

  2. The two-spread-view effect is positive on LIDC but negative on LCTSC despite the analogous acquisition layout. Do existing diagnostics identify whether the difference is associated with population alignment, scan extent, or fitting behavior? I am asking for an interpretation grounded in available results, rather than a new multi-backbone study during rebuttal.

  3. Please specify the exact calibrated weight function wjuw_{ju} in Eq. (1), including its support, the empty-support clamp, and where the stated temperature schedule enters it. A short formula or implementation would make the central fixed-lifting baseline reproducible.

  4. Which of the proposed mechanisms and remedies were prespecified before inspecting the reused LIDC/TotalSegmentator assessments, and which are exploratory follow-ups? A concise table of primary locked tests versus development analyses would help my evaluation.

2 comments
AuthorsOpenReview ↗

Response to Reviewer K6kX

Q: The results establish pipeline-specific coordinate utility more strongly than a general mechanism. FULL also fits twelve-view training cases worse than OFF (Table 15), which leaves optimization and fusion interference as alternative contributors. Given this, what do the existing diagnostics distinguish between measurement-dependent positional bias and optimization/interference introduced by the additive coordinate pathway? Which mechanistic conclusion is supported beyond the fixed-budget, pipeline-specific effect?

A: We agree that the diagnostics do not uniquely identify the cause of the twelve-view penalty. They establish dependencies on training conditions and model design, rather than a general mechanism. Section 5.1 (Table 2) and Appendix I.1 report the extended-budget comparison. At 11,250 updates, the three-seed mean FULL−OFF differences are:

CohortMixed-view: trainingMixed-view : evaluationTwelve-view-only: evaluation
LCTSC+0.51−11.28−0.10
LIDC-IDRI−0.10−1.43−0.02

All values are Dice percentage points. The mixed-view evaluation penalty persists despite small training-fit differences. However, neither these differences nor the longer budget establish convergence or rule out optimization effects.

Twelve-view-only training substantially reduces the penalty, but also provides five times as many twelve-view examples and 2.22 times as many projections at equal update counts. When we match twelve-view exposure instead of updates (Appendix I.1), the regime difference remains on LCTSC (+11.45 points, interval 8.46 to 14.71) but not on LIDC-IDRI (-0.57, interval -1.30 to 0.25), so exposure alone does not explain the LCTSC contrast. Training-distribution effects and optimization/generalization dynamics therefore remain unresolved.

The single-seed concatenation control also exhibits the mixed-view penalty, so it is not unique to additive injection in the tested controls. Conversely, the DIF-Net adaptation does not reproduce the reversal (Section 5.4). The directional and context controls in Section 5.2 constrain possible explanations but do not identify the cause of the broad-view penalty.

Our supported conclusion is therefore conditional: the coordinate penalty persists under longer mixed-view training in the main pipeline, but changes with training regime and architecture. We do not claim that measurement-dependent positional bias has been isolated from optimization or fusion interference.

Q: The explanatory interventions have uneven replication. The full context-by-coordinate interaction is single-seed and the directional analyses remain on reused development cohorts. The FULL−OFF contrast is an end-to-end comparison between separately trained systems, not a direct estimate of a coordinate signal’s intrinsic effect. The conclusion should remain specific to the tested pipeline and training protocol.

A: We agree that FULL−OFF compares separately trained systems, not the intrinsic effect of a coordinate signal. It measures the end-to-end effect of including coordinates under the specified architecture and training protocol, including any changes to optimization and the learned solution.

  • Section 5.2, Table 4 now reports the complete context-by-coordinate comparison across three paired seeds. Section 5.3 also reports the three-seed capacity-matched slot control; this is distinct from the earlier single-seed component factorial.

  • The directional, context, and slot follow-ups still use previously inspected LIDC-IDRI development patients. Additional seeds assess training variability but do not provide an independent cohort confirmation. The concatenation and DIF-Net adaptation controls remain single-seed.

We therefore restrict the conclusion to the tested pipelines and training conditions. These interventions establish when coordinate utility changes, but do not isolate a general causal mechanism for the broad-view penalty.

AuthorsOpenReview ↗

Response to Reviewer K6kX

Q: The two-spread-view effect is positive on LIDC but negative on LCTSC despite the analogous acquisition layout. Do existing diagnostics identify whether the difference is associated with population alignment, scan extent, or fitting behavior?

A: Figure 3 and Section 5.1 report the original cross-cohort disagreement explicitly. We also examined existing checkpoints, acquisition geometry, and anatomical measurements. These diagnostics constrain the interpretation but do not identify the cause.

At 11,250 updates, the three-seed mean FULL−OFF differences for the spread pair [0°, 90°] are:

CohortTraining patientsEvaluation patients
LCTSC+1.78−8.33
LIDC-IDRI+1.51+0.02

Values are Dice percentage points. These extended-budget follow-ups are distinct from the original Figure 3 results; Appendix I.3 (Table 18) gives the full comparison. The LCTSC penalty is absent on its training patients, which is consistent with a larger generalization gap for FULL rather than poorer training fit at this checkpoint. It does not establish the underlying mechanism.

The geometry audit confirms that both cohorts use full-volume bounds, not organ-mask crops. Cohort size, label source, anatomical distribution, and acquisition details still differ. Exploratory associations with scan extent, lung position, and detector spacing do not isolate any of these as the cause.

We therefore leave the cross-cohort difference unresolved and do not attribute it specifically to population alignment, scan extent, or optimization.

Q: Please specify the exact calibrated weight function in Eq. (1), including its support, the empty-support clamp, and where the stated temperature schedule enters it. A short formula or implementation would make the central fixed-lifting baseline reproducible.

A: Section 3 now explains Eq. (1) term by term: the weight w_ju, its support (the set of image features allowed to contribute to a query), and the clamp applied to the denominator when the support is empty. Appendix A.3 states that the routing temperature decreases from 4 to 1 over the presentation horizon. Section 3 now also gives the weight explicitly (Equation 2). For each view, the query point is projected onto the 16×16 feature grid, and the 5×5 cells around the nearest cell centre are candidates. Each candidate receives a Gaussian weight of its distance to the projected point, with width 0.6 cells, normalized over the 25 candidates (denominator clamp 10^-12). Candidates with an invalid projection, outside the grid, or with a normalized weight at or below 10^-6 receive zero weight; the remaining ones form the support. The temperature T enters as an exponent 1/T on the retained weights. It follows a cosine schedule from 4 to 1 during training and equals 1 at evaluation; Appendix A.3 gives the horizon for each cohort. The pooled feature in Eq. (1) divides by the sum of the weights over all views, clamped at 10^-8, and a query with empty support receives a zero feature before layer normalization. We will release the code upon acceptance.

Q: Which of the proposed mechanisms and remedies were prespecified before inspecting the reused LIDC/TotalSegmentator assessments, and which are exploratory follow-ups? A concise table of primary locked tests versus development analyses would help.

A: The information is in the revised text, and we have added the table you ask for (Table 22, Appendix I.7).

  • Locked test: the LCTSC OFF/FULL comparison. Its protocol, contrasts and checkpoint rule were fixed before fitting, and all six models were frozen before test prediction (Section 4.2, Appendix G.2).

  • Development analyses: the gate, directional, aggregation, context and synthetic results. Section 4.2 states that locking a later experiment does not make these assessments fresh tests.

  • Exploratory follow-ups: the context variants, broader two-view training and fusion controls (Appendix F; the fusion controls are summarized in Appendix D.3), and the earlier remedies listed in Appendix H.

  • Rebuttal follow-ups: the longer-training, twelve-view-only, equal-exposure, mean-mask, coordinate-only, wider-context, directional-anchor, capacity-matched slot, concatenation, DIF-Net and fastMRI controls reported in this letter. Except for the new fastMRI assessment, they reuse previously inspected patients.

  • Table 22 lists these four groups with the data each used. It also records that the gate cutoff of 0.25 was frozen before the twelve-run LIDC-IDRI comparison and kept unchanged for TotalSegmentator, and that its initial choice is not documented.

Reviewer_NPKkOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper is intended to incorporate projections, namely digitally reconstructed radiographs for CT lung segmentation. The contributions are not clearly stated. The experiments were performed on a pulic CT dataset named LIDC-IDRI.

Critical strengths

None

Critical weaknesses
  • The problem background and methodological formulation of this paper are unclear. Citations do not match with the description.
  • The method and results are poorly presented. Throughout the paper, no projection and segmentation results were visualized.
Questions for authors
  1. In the abstract and introduction sections, the authors did not formulate the background clearly. For example, in the sentence "a calibrated X-ray projection associates a 3D query with a detector location", what is a calibrated X-ray projection, what 3D query means, and what detection location represents. The authors should clarify the problem this paper is addressing clearly so that readers can understand.

  2. Citations in the main text do not match with the description. For example, "... with anatomical regularities to resolve this ambiguity (Arridge et al., 2019)." I check this reference and it is not related to anatomical regularities. The authors should seriously resolve these issues.

  3. This paper addresses the problem of lung segmentation from CT. However, the paper does not visualize any segmentation and CT projections. Besides, the authors should clarify how the projections are generated from CT. The authors should address these issues to improve the quality of this paper.

1 comment
AuthorsOpenReview ↗

Response to Reviewer NPKk

Q: The problem background and methodological formulation are unclear, and the contributions are not clearly stated. For example, in the sentence “a calibrated X-ray projection associates a 3D query with a detector location”, what is a calibrated X-ray projection, what does 3D query mean, and what does detector location represent?

A: We rewrote the abstract and introduction so that each of these terms is defined before it is used.

  • Section 1, second paragraph: states the task (predict a 3D mask of the left and right lungs from one or more simulated X-ray images) and that the output is a segmentation mask, not a reconstructed CT image.

  • Section 1, same paragraph: “calibrated” means the positions of the X-ray source and detector relative to the patient volume are known; the detector location is where the ray through a 3D point meets the detector; a “3D query point” is a point at which the model gathers image features.

  • Section 1 now ends with three explicit contributions, each with a section reference.

  • Section 4.1 lists all three datasets (LIDC-IDRI, TotalSegmentator, LCTSC) with their splits and label sources. The held-out test result is on LCTSC, not LIDC-IDRI.

Q: Citations in the main text do not match the description. For example, “... with anatomical regularities to resolve this ambiguity (Arridge et al., 2019)”. This reference is not related to anatomical regularities.

A: Thank you for identifying this mismatch. We removed the sentence linking Arridge et al. (2019) to anatomical regularities. Section 2 now cites it only for combining physical measurement models with learned priors, which the review discusses. We no longer use this reference to support an anatomy-specific claim. We also checked the remaining citations against their sources.

Q: The paper does not visualize any segmentations or CT projections. The authors should also clarify how the projections are generated from CT.

A: We have added both the rendering description and visual examples.

  • Section 3 explains that we convert CT values to attenuation and integrate along rays from a simulated point source to detector pixels. The source-to-centre and source-to-detector distances are 1500 and 1800 mm. Section 4.1 describes the 24 projections at 7.5° intervals, their 512² resolution, and preprocessing. Appendix A.1 gives the conversion formula and the ray integration (256 samples per ray), and clarifies that the simulations do not reproduce scatter, beam hardening, or scanner-specific noise.

  • Figure 4, discussed in Section 5.1, shows input DRRs and OFF/FULL lung predictions against ground-truth outlines at one and twelve views, alongside mean-mask and coordinate-only references. Its caption now specifies the selection rule: the median and lowest-scoring patients within a three-patient pool (one patient per institution) that was fixed before the follow-up training runs, ranked by mean Dice across OFF/FULL and one/twelve-view conditions. Predictions use mixed-view models with seed 0 at 11,250 updates, and the displayed scores are whole-volume 3D Dice.

The figure provides both a typical example within the fixed subset and a lower-performing example. It does not include two-view layouts or error maps.

Reviewer_UgMgOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper explores additional geometry information in general projection-to-volume reconstruction task, focusing on the multiple-view X-ray to 3D lung segmentation problem. It uses three public datasets, LIDC-IDRI, TotalSegmentator and the LCTSC dataset. The authors explore whether, in a calibrated setup where the patient's position relative to the detector and source and the overall geometry are known, this additional information can be used to disambiguate and improve 3D anatomical segmentation from multiple views. They find that views that are very near each other benefit, but when the views are far apart they do not benefit and can even harm performance.

Critical strengths

The problem being considered is an interesting one where there are multiple ways to aggregate multi-view x-ray to 3D reconstruction and providing disambiguating information from extrinsic geometry parameters is an interesting take.

Critical weaknesses

The paper is very dense and hard to read: non-standard terms such as "intermediate angle deficit" (introduced in the abstract), "calibrated transport", "location-indexed learned embedding" etc. are introduced without explanation and has to be understood from the context. The fig 2 describes the overall pipeline but does not leverage the visual medium to provide intuition and just brings the equations from the text into the diagram. There is a "Equivariance and symmetry breaking" section in the related work section that seem very far-off from the problem being addressed. The context setup is not clearly described.

experiments and results do not seem rigorous: The main text points to table 4 (which is in the appendix section) seem to suggest that better decoder (kernel dilation with wider context) improves dice score significantly (0.48 to 0.63) without the additional 3d coordinate information. The use of "local decoder" (with kernel dilation of 1) in the main results seem to inflate the improvement.

Questions for authors
  • (S) A more visual and interpretative diagram would have helped in figure 2. A ray-geometry diagram explaining the problem setup and describing the 3D coordinates being added would have helped the reader.
  • (S) the main task is multi-view X-ray to 3D lung segmentation, but the paper then turns to purely synthetic 2D toy problems (sparse k-space sampling, parallel-beam projection). Could real data such as MRI k-space have been used?
4 comments
AuthorsOpenReview ↗

Response to Reviewer UgMg

Q: The paper is very dense and hard to read: non-standard terms such as “intermediate angle deficit” (introduced in the abstract), “calibrated transport”, “location-indexed learned embedding” etc. are introduced without explanation and have to be understood from the context.

A: We rewrote the abstract and Sections 1–5 in plain language.

  • The three terms you list no longer appear. “Intermediate angle deficit” is now “fails at intermediate angular separations”; “calibrated transport” is now “geometry-based feature sampling”; “location-indexed learned embedding” is now “spatial slots”, defined in Section 5.3 and Appendix A.5 as learned feature vectors stored at each 3D grid location.

  • Section 1 opens with the task, its input and its output, and defines “calibrated” and “3D query point” before using them.

  • Section 4.2 defines the acquisition terms (layout, clustered, spread, offset, held) at first use.

Q: Fig. 2 describes the overall pipeline but does not leverage the visual medium to provide intuition and just brings the equations from the text into the diagram. A ray-geometry diagram explaining the problem setup and describing the 3D coordinates being added would have helped the reader.

A: We agree and have replaced the figure. The new Figure 2 has two panels. Panel (a) is a ray-geometry diagram: it shows the X-ray source, the patient volume with its grid of query points, the detector, and one ray through a query point. It shows that all points on that ray share one detector location, illustrating why detector location alone cannot distinguish positions along a ray. Panel (b) shows the pipeline. The figure separates geometry-based sampling, which remains active in every model, from the coordinate input, and its legend identifies OFF, FULL and SCALAR. We also added Figure 1 in the Introduction, a small plot of lung Dice against the number of views on the LCTSC test patients, with and without coordinates, so that the main result can be read at a glance.

Q: There is an “Equivariance and symmetry breaking” section in the related work that seems very far-off from the problem being addressed.

A: We reduced this material to two sentences in the last paragraph of Section 2 and tied it to its one use in the paper. They say that the cited results motivate the synthetic translation experiments and nothing else, and point to Appendix B for the assumptions and for why the argument does not describe the cone-beam pipeline.

Q: The context setup is not clearly described.

A: We added a paragraph on decoder context to Section 3. It defines decoder context as the neighbourhood of 3D features a convolution can use, and states the exact change: the two convolutions at the 32³ stage go from dilation 1 in both layers to dilation 2 and 3, with matching padding, while initial weights, parameter count, and the coordinate-branch stay unchanged. Appendix A.3 repeats these settings.

Q: Experiments and results do not seem rigorous: the main text points to Table 4 (which is in the appendix) and suggests that a better decoder (kernel dilation with wider context) improves Dice significantly (0.48 to 0.63) without the additional 3D coordinate information. The use of a “local decoder” in the main results seems to inflate the improvement.

A: We agree that the local decoder makes the one-view coordinate benefit larger. Section 5.2, Table 4 now reports the complete local/wider-context × OFF/FULL comparison across three paired seeds on 37 LIDC-IDRI development patients, at 11,250 updates.

  • Wider context raises one-view OFF Dice from 53.47% to 60.72%, while FULL changes from 80.37% to

79.96%. The coordinate advantage therefore falls from 26.90 to 19.24 Dice points, a reduction of 7.66 points.

  • At twelve views, wider context reduces the average coordinate penalty from −1.43 to −0.38 points.

Decoder context therefore explains part of the measured coordinate advantage, but a substantial one-view benefit remains. We now state this dependence explicitly rather than treating the effect as independent of decoder design. The separate coordinate-free experiment in Appendix F.1 uses a different protocol and is reported separately.

AuthorsOpenReview ↗

Response to Reviewer UgMg

Q: The main task is multi-view X-ray to 3D lung segmentation, but the paper then turns to purely synthetic 2D toy problems (sparse k-space sampling, parallel-beam projection). Could real data such as MRI k-space have been used?

A: Yes. We have added a real-data fastMRI knee reconstruction experiment in Section 5.4, with the full table in Appendix I.6 (Table 21). We retrospectively undersample acquired multi-coil k-space at R=4 and R=8 and train matched U-Nets with and without image coordinates. Both use the same architecture, initialization, masks, and data order within each of three paired seeds. Assessment covers 100 volumes from 96 patients.

Mean assessment SSIM:

AccelerationOFFFULLFULL−OFFSeed SD of difference
R=40.9357780.936216+0.0004380.000419
R=80.8995190.900209+0.0006890.000763

Coordinates yield small average improvements, with appreciable seed variation. Appendix I.6 also reports paired-patient bootstrap intervals conditional on these fitted seeds. We present this as an exploratory real-data extension, not a replication of the CT mechanism.

We retain the synthetic experiments because they permit controlled changes to population alignment, measured frequencies, and decoder context (Sections 4.1 and 5.2). For example, we can translate objects while keeping the measurement operator and sampling budget fixed and not the measured signals themselves. Under periodic Fourier sampling, the large coordinate benefit for aligned objects disappears after random translation. These controlled diagnostics and the real-data MRI experiment therefore serve different purposes.

I would like to thank the authors for the effort and engagement. Regarding the readability issue, there has been an effort to simplify the non-standard terms and addition of figures for context setup. The extensive rewrite now means most of the main text is either blue or red. On the other hand, i think the readability concern is only partly addressed. For example, the main text is supposed to stand on its own but there are too many references back to the appendices. There are texts and sentences which seem to not follow serially and are out of place (for example results inside the setup section), making arguments hard to follow.

Regarding the experiment results, i agree with fellow reviewers that findings are narrow.

AuthorsOpenReview ↗

Follow-up Response to Reviewer UgMg

  1. Thank you for rereading the paper. Your comments helped us see what the rewrite missed. And we tried to simplify the terminology, but the order of the argument still makes the reader do too much work. We will put the information needed to the main part, so the appendix is optional for following the argument.

  2. We agree that the finding is specific to the tested pipeline. The practical point we want readers to take away is clearest at two views: on the same held-out LCTSC patients, adding coordinates improved Dice by 19.55 points when the views were 15° apart, but reduced it by 3.63 points when they were 90° apart. The trained models were the same in both evaluations. In this setting, reporting only “two views” would hide a difference large enough to reverse the recommendation.

Reviewer_6xZmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper does not propose a new reconstruction network. It studies when explicit coordinates help within one fixed projection-to-volume pipeline. Calibrated X-ray features are lifted to a 16³ query grid, and a 3D decoder outputs a 96³ left/right lung segmentation. The paper compares no coordinates (OFF), a coordinate MLP (FULL), and a scalar gate that scales the coordinates according to acquisition geometry (SCALAR). All projections are rendered from CT; no measured radiographs are used. The main result is that the sign of the coordinate benefit depends on measurement coverage. On a prospectively locked, previously unused LCTSC test set (24 patients), FULL improves over OFF by 30.81 Dice points at one view but loses 6.34 points at twelve views. With two views, clustered layouts gain 19.55 points while spread layouts lose 3.63. The paper also compares along-ray depth with detector-plane coordinates, varies decoder context, and studies population alignment on synthetic data. I view the contribution as a carefully controlled empirical study rather than a new method or performance result.

Critical strengths

1.The locked fresh cohort validates a non-trivial reversal. LCTSC had not been used in the project, all six models were frozen before testing, and the criteria were declared in advance. Table 3 reports +30.81 at one view (95% interval [28.96, 32.38]) and −6.34 at twelve views ([−7.86, −4.90]), with consistent signs across three initializations. At the same view count, clustered and spread layouts show opposite effects (+19.55 versus −3.63), so view count alone does not predict whether coordinates help. 2.The attribution design and reporting of negative results are careful. Calibration is retained in every arm, while coordinates, aggregation, context, and coordinate direction are controlled separately. The authors fully report unfavorable results, including gate failure at intermediate spans, remedies that miss their prespecified screens, and worse training fit at twelve views. Tables 7 and 8 list all acquisition layouts rather than only the passing subset.

Critical weaknesses

1.The independently validated scope is narrow, and the cause of the twelve-view penalty is not isolated. LCTSC validates only the OFF/FULL coverage interaction; the depth/plane, context, slot, and gate conclusions are evaluated only on repeatedly reused development cohorts, as the authors acknowledge in Section 7 and Appendix I. All projections are simulated DRRs, and the clinical evidence comes from one model with roughly 1.31M active parameters, one additive coordinate injection, and one coarse lung-segmentation task, with no measured radiographs or second architecture. Training code, checkpoints, and raw predictions are not released, so the results cannot be independently audited. Even the confirmed reversal is confounded at the dense end. At twelve views, FULL also fits the training set worse than OFF (0.8381 versus 0.8585, Table 15), while each LCTSC model receives only 2,250 updates compared with 11,250 on LIDC. The −6.34 result may therefore partly reflect underfitting or optimization, rather than coordinates being harmful under broad coverage. The title and abstract should be narrowed to this calibrated DRR-to-lung-segmentation pipeline. 2.The gate descriptor is misaligned with the measured utility of coordinates. Eq. (3) uses a cutoff of 0.25, and Appendix D gives λ_min(H) = (1−cos θ)/2 for coplanar two-view geometry, so the gate is zero for separations of 60° or more. Yet Table 7 reports FULL−OFF of +9.34 at 0°/60° and still +4.28 at 0°/75°: coordinates remain useful in these layouts, but the gate switches them off. Consequently, at 0°/60° SCALAR (0.7614) nearly equals OFF (0.7601) and trails FULL (0.8535) by about nine points. Broader two-view training raises SCALAR to 0.7996, still below the prespecified 0.82 target, and three primary retention contrasts still fail (Appendix F.2). The authors attribute this to the descriptor, and I agree. This means, however, that the gate, one of the four stated contributions, holds only at the prespecified endpoints and should not be presented as a solution to the sparse/broad trade-off. The paper states that the 0.25 cutoff was fixed before this campaign; since the same development cohort had been used repeatedly, please clarify whether earlier results informed this choice.

Questions for authors

1.Can you report a population-mean or atlas-mask baseline? LCTSC single-view FULL reaches only 0.6719; such a baseline would show whether coordinates mainly supply a mean-shape prior or help exploit projection evidence. 2.Does the twelve-view penalty persist with LIDC-level update counts, concatenation-based injection, or a wider decoder? Can the OFF/FULL reversal be reproduced in an existing projection-to-volume architecture? 3.Can FULL and OFF arms be added to the directional replication? Without them, a depth benefit cannot be distinguished from degradation of the detector-plane arm; the much larger seed SD of the plane arm in Table 2 (0.0282 versus 0.0008) points to the latter possibility. The slot ablation also removes 1,048,576 active parameters; can it be repeated with capacity-matched controls and multiple seeds?

5 comments
AuthorsOpenReview ↗

Response to Reviewer 6xZm

Q: The independently validated scope is narrow. LCTSC validates only the OFF/FULL coverage interaction; the depth/plane, context, slot, and gate conclusions are evaluated only on repeatedly reused development cohorts. All projections are simulated DRRs, and the clinical evidence comes from one model, one additive coordinate injection, and one coarse lung-segmentation task. Training code, checkpoints, and raw predictions are not released. The title and abstract should be narrowed to this calibrated DRR-to-lung-segmentation pipeline.

A: We agree with this description of the scope and have narrowed the paper to match it.

  • The abstract and Section 1 now state that the study concerns one projection-to-volume pipeline, simulated X-ray projections rendered from CT, and a lung-mask output. The contribution list says “within this pipeline”.

  • Section 6 (Limitations; Section 7 in the original) states which results rest on the held-out LCTSC test (OFF/FULL only) and which rest on reused development cohorts. Appendix H records the earlier exposure of each cohort.

  • The title is now “When Do Coordinates Help? A Controlled Study of Simulated X-ray to Lung-Mask Prediction”.

  • We will release the code upon acceptance. The Reproducibility Statement now states this.

Q: The cause of the twelve-view penalty is not isolated. At twelve views, FULL also fits the training set worse than OFF (0.8381 versus 0.8585, Table 15), while each LCTSC model receives only 2,250 updates compared with 11,250 on LIDC. The −6.34 result may therefore partly reflect underfitting or optimization, rather than coordinates being harmful under broad coverage.

A: We ran the control you describe. We trained OFF and FULL on LCTSC and LIDC-IDRI for 11,250 optimizer updates, with checkpoints at 2,250 and 11,250, under mixed-view and twelve-view-only training. Results are averaged across three paired seeds (0, 1 and 2).

  • Longer training does not remove the mixed-view penalty. At 11,250 updates, the twelve-view FULL−OFF difference is −11.28 Dice percentage points on LCTSC test patients and −1.43 on LIDC-IDRI development patients, with negative differences in every seed. The corresponding training-set differences are small (+0.51 and −0.10 points, respectively). This does not rule out optimization effects.

  • Extending the budget does not resolve the concern by itself. On LCTSC, the average penalty increases from −6.12 points at 2,250 updates to −11.28 at 11,250. This shows that the penalty persists with five times more updates; it does not establish convergence or rule out optimization effects.

  • Under twelve-view-only training, the average difference is near zero: −0.10 points on LCTSC and −0.02 on LIDC-IDRI. Patient-bootstrap intervals include zero on both cohorts, although this does not establish equivalence between OFF and FULL.

These results support a narrower conclusion: the twelve-view coordinate penalty depends on the training setup in this pipeline. At equal update counts, twelve-view-only training provides five times as many twelve-view examples and 2.22 times as many total projections. The comparison therefore does not separate training-distribution effects from increased exposure or optimization dynamics. All follow-ups reuse previously inspected patients and are not new independent confirmations. Patient-bootstrap intervals describe uncertainty across patients conditional on the fitted seeds; variation across seeds is reported separately. Section 5.1 (Table 2) and Appendix I.1 report these results.

AuthorsOpenReview ↗

Response to Reviewer 6xZm

Q: The gate descriptor is misaligned with the measured utility of coordinates. With cutoff 0.25 the gate is zero for separations of 60° or more, yet Table 7 reports FULL−OFF of +9.34 at 0°/60° and +4.28 at 0°/75°. The gate, one of the four stated contributions, holds only at the prespecified endpoints and should not be presented as a solution to the sparse/broad trade-off. Please clarify whether earlier results informed the choice of the 0.25 cutoff.

A: We agree. The gate switches coordinates off in layouts where they still help, and the paper no longer presents it as a solution to the sparse/broad trade-off.

  • The contribution list now has three items. The third, “A test of geometry-based gating,” explicitly acknowledges failure at intermediate angular separations.

  • The gate now has one paragraph in Section 5.3; its definition is in Section 3 (Equation 4) and Appendix D, and its full results are in Appendix D. The text reports the 0°/60° result (SCALAR 0.761, FULL 0.853), the failed remedy checks after broader training, describes the gate as a baseline for selected endpoints and not as a solution, and states that the 0°/60° pair has a zero gate, so the rule removes the coordinate input in a layout where FULL outperforms OFF, and that SCALAR empirically scores close to OFF there.

  • Appendix D.1 states that the cutoff is a hyperparameter and that H is not a measure of missing volumetric information. Appendix D gives the two-view formula and notes that the gate reaches zero at 60°.

  • The cutoff of 0.25 was a heuristic choice made during development, not a theoretically derived constant. It was locked before the twelve-run comparison, was not searched or adjusted using that campaign’s outcomes, and was retained unchanged for the TotalSegmentator replication. How it was first chosen is not documented, so we cannot exclude that earlier results on the same development patients informed it. Appendix D.1 and Appendix I.7 now state this.

Q: Can you report a population-mean or atlas-mask baseline? LCTSC single-view FULL reaches only

0.6719; such a baseline would show whether coordinates mainly supply a mean-shape prior or help exploit projection evidence.

A: We added two image-free baselines: a population-mean mask computed from training labels only, and a coordinate-only model. The mean mask predicts the same training-derived anatomy for every patient. Neither baseline uses projection images, so its predictions do not change with view count.

Model/inputLCTSC testLIDC-IDRI development
Training-set mean mask55.3772.59
Coordinate-only57.4675.12
OFF, one view38.9053.47
FULL, one view68.1680.37
OFF, twelve views83.8292.85
FULL, twelve views72.5391.42
  • Matched comparison. The table reports mean Dice (%) at 11,250 updates, averaged across three seeds for trained models. OFF and FULL use mixed-view training. FULL exceeds both image-free baselines. Its long-budget performance is therefore not explained by a fixed population-shape prediction alone. This comparison does not quantify its reliance on image evidence.

  • Anatomical priors remain important. One-view OFF performs below both image-free baselines. However, this does not establish what fraction of the OFF-to-FULL improvement comes from a shape prior.

  • The conclusion depends on training budget. At 2,250 updates, LIDC one-view FULL scores 68.99%, below the mean mask (72.59%) and coordinate-only model (72.63%). We retain this result to avoid implying that FULL always exceeds anatomy-only predictions. Section 5.1 reports these references, Figure 4 shows them for two test patients, and Appendix I.2 (Table 17) gives the full table.

AuthorsOpenReview ↗

Response to Reviewer 6xZm

Q: Does the twelve-view penalty persist with LIDC-level update counts, concatenation-based injection, or a wider decoder? Can the OFF/FULL reversal be reproduced in an existing projection-to-volume architecture?

A: We completed all four controls. Below, differences are FULL−OFF in Dice percentage points at 11,250 updates.

  1. Longer training. Section 5.1 (“View coverage”) reports three paired seeds on both cohorts. Under mixed-view training, the twelve-view penalty persists: −11.28 points on LCTSC and −1.43 on LIDC-IDRI, negative in every seed. Under twelve-view-only training, the differences are near zero (−0.10 and −0.02). This comparison changes both the training distribution and broad-view exposure, so it does not isolate mixed-view interference.

  2. Concatenation-based injection. On LIDC-IDRI, the twelve-view difference is −1.70 points under mixed-view training and +0.07 under twelve-view-only training. Thus, the qualitative pattern also appears with concatenation. OFF and FULL have matched parameter counts within this control. Section 5.4 and Appendix I.6 report it.

  3. Wider decoder. Section 5.2 (“What changes the effect”) now reports the completed three-seed comparison. Wider context reduces the twelve-view mixed-training penalty from −1.43 to −0.38 points. The average remains negative, but its magnitude depends on decoder context.

  4. Existing architecture. Section 5.4 (“Extensions beyond the main pipeline”) reports our DIF-Net adaptation for lung segmentation. It keeps DIF-Net's 2D encoder, geometry-based feature sampling and per-point MLP decoder, takes the maximum over views in each feature channel, and replaces the intensity output with three class scores (background, left lung, right lung) trained with cross-entropy plus a soft lung Dice loss. FULL adds three patient-centered coordinate channels to the decoder input and OFF supplies zeros in those channels. It is a DIF-Net-based segmentation adaptation, not an unchanged reproduction of the original reconstruction model (Appendix I.6). The reversal does not reproduce: coordinates improve Dice by 42.34 points at one view and 2.87 at twelve views on LIDC-IDRI.

The concatenation and DIF-Net comparisons use seed 0 only, and these follow-ups reuse previously inspected patients. Together, the results support restricting the reversal claim to the tested pipeline and training conditions, rather than treating it as a general property of projection-to-volume models.

Q: Can FULL and OFF arms be added to the directional replication? Without them, a depth benefit cannot be distinguished from degradation of the detector-plane arm; the larger seed SD of the plane arm (0.0282 versus 0.0008) points to the latter. The slot ablation also removes 1,048,576 active parameters; can it be repeated with capacity-matched controls and multiple seeds?

A: We have completed both requested controls on LIDC-IDRI across three matched seeds and updated Sections 5.2 and 5.3.

  1. Directional anchors (Section 5.2). We added OFF and a full acquisition-coordinate arm to the depth-only and detector-plane comparison. At source acquisitions, their gains over OFF are +22.93, +13.25, and +23.42 Dice points, respectively. The patient-bootstrap intervals for depth−OFF and plane−OFF are [21.62,

24.14] and [11.93, 14.49]. Both components therefore improve over OFF: depth’s advantage is not simply detector-plane inputs harming performance. The same ordering holds under held rotations. Plane inputs remain more seed-sensitive, and differing active coordinate dimensions prevent a depth-specific mechanistic conclusion. Appendix I.4 (Table 19) gives all contrasts.

  1. Capacity-matched slots (Section 5.3, Appendix I.8). We compared spatial slots, no slots, and a shared non-spatial module with 1,048,576 active parameters, matching the removed slot parameters, on LIDC-IDRI (178 training patients and 37 reused development patients) with three matched seeds. Explicit coordinates remain enabled in all three arms. At 11,250 updates, slots improve one-view Dice over no slots by ~0.62 points, while the shared module improves it by ~0.41 points. Additional non-spatial capacity thus recovers much of the improvement, weakening attribution to spatial slots alone. One-view Dice is 82.17 [80.11, 83.99] with slots, 81.55 [79.60, 83.25] without slots and 81.95 [79.94, 83.71] with the shared module. Paired 95% patient-bootstrap intervals for slots minus the shared module include zero in 15 of 16 comparisons: all eight source layouts and seven of the eight offset layouts. Only the offset one-view interval excludes zero (+0.27 [+0.04, +0.51]). No multiplicity adjustment is applied. These results do not establish equivalence between the two modules. The full results are in the new Table 23 (Appendix I.8).

These follow-ups reuse development patients. Patient uncertainty is reported conditional on the fitted seeds, with training-seed variation reported separately.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Reviewers and Summary of Revisions

Dear Reviewers and Area Chair,

We thank you for your careful reading and your comments, which helped us improve the paper. We have rewritten the manuscript for clarity, narrowed its claims to the tested pipeline, and run the controls you asked for. These controls changed what the paper claims: the broad-view penalty persists with longer mixed-view training, nearly disappears under twelve-view-only training, and does not appear in the DIF-Net segmentation adaptation. The revised claim is therefore narrower and better supported: whether coordinates cost accuracy at broad coverage depends on the training regime and the architecture. Each comment is answered below with the evidence, the remaining limitation, and the change made to the manuscript.

Note:

  • Any information modified in the paper as a response to reviewers’ comments is presented in blue color in the revised paper.

  • Any new information added to the paper as a response to reviewers’ comments is presented in red color in the revised paper.

  • Section, table and figure numbers in this letter refer to the revised paper. Where a reviewer cited the original numbering, we give both.

  • Except for the fastMRI experiment, which uses new data, the follow-up controls reported here reuse the LCTSC test patients and the LIDC-IDRI development patients. They are follow-up analyses, not new untouched tests. The follow-up models were trained again from scratch, so their values at a matching budget differ from the original runs; Appendix I.1 explains the differences, and Section 4.2 states which rotations a “source” value averages.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.