acceptodds
Withdrawn from ICLR 2027

AtlasS2V: Organ Atlas–Guided Prediction of 3D Tissue Distributions from a Single Section

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Three-dimensional tissue reconstruction typically relies on multiple spatially pro- filed sections from the same specimen, limiting its applicability to valuable datasets represented by a single section. We introduce ATLASS2V, which combines one observed 2D tissue section with a spatially registered 3D reference atlas to pre- dict cell-density and cell-type distributions above and below the section. The atlas supplies depth-resolved anatomical structure and spatial context beyond the observed plane, while the observed section captures specimen-specific devi- ations from this population prior. ATLASS2V calibrates atlas abundance to the observed section and transfers its residual through a reference-trained distance gate, retaining more individual information nearby and relying increasingly on the atlas at greater distances. On an independent mouse, ATLASS2V improves reconstruction over central-copy and Atlas-only across near-to-intermediate offsets and converges toward the atlas at longer distances. Strict outer-animal validation shows consistent gains over central-copy in all four held-out mice, increasing mean composition Pearson from 0.392 to 0.629. A marker-density adaptation further improves distance-integrated correlation over each animal’s best fixed expert in 58 of 60 held-out mice. Organ-specific studies demonstrate feasibility in human lung, breast, and kidney. Atlas-derived context also improves sparse-panel inference: with ten measured genes, hidden-gene correlation increases from 0.6755 to 0.7268, while 34-class cell-typing macro-F1 increases from 0.3058 with 2D atlas context to 0.4707 with correct 3D context. These results support reference-conditioned prediction as a route to extending single-section datasets into virtual 3D tissue distributions without collecting serial sections from each target specimen.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Spatial transcriptomics, 3D tissue reconstruction, single-section reconstruction, reference atlas, cell-type distribution, cell-density prediction, spatial omics, virtual tissue reconstruction, atlas-guided modeling, histology-to-3D prediction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_yY5ROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ATLAS S2V, a reference-conditioned method for estimating 3D cell-density and cell-type distributions from one observed 2D tissue section and a spatially registered organ atlas. It calibrates atlas abundances to the observed specimen, forms a spatial residual relative to the reference, and transfers that residual with a distance-dependent two-expert gate. The authors report brain reconstruction on held-out mice, organ-specific extensions to lung, breast, and kidney, and separate atlas-conditioned readouts for sparse-panel gene imputation and cell typing. The output consists of virtual, grid-based distribution estimates rather than the actual positions of unseen cells.

Critical strengths
  1. Predicting unobserved tissue distributions when only one target section is available is relevant to valuable or archived spatial-omics specimens. An external anatomical atlas is a sensible prior for this underdetermined problem, and the paper acknowledges that exact unseen cells cannot be recovered. 2.The abundance calibration, residual correction, and distance-aware mixture are simple and understandable. The paper reports comparisons with central-copy, atlas-only, nearest-reference, and U-Net alternatives, as well as held-out-animal and multiple-organ experiments. The sparse-panel analyses also offer evidence that anatomical atlas context can be helpful.
Critical weaknesses
  1. The added specimen-specific 3D information beyond the atlas is not convincingly isolated. The results suggest that the reference atlas may explain much of the depth-dependent structure. In Table 1, the class Pearson of ATLAS S2V and atlas-only is identical at 2,000 micrometers (0.553) and nearly identical at 800 micrometers (0.539 versus 0.538). In Table 3, the full model achieves 0.5983, compared with 0.5737 without residuals and 0.5953 with fixed 1:1 fusion. Although the 60-mouse density study reports a stronger routing benefit, the primary reconstruction results leave the individual-specific contribution and necessity of the proposed gate uncertain. The 3D visualization in Fig. 2 does not clearly separate atlas-derived anatomy from newly inferred individual structure. The manuscript should quantify the incremental spatial information, provide atlas-minus-prediction and prediction-error maps, and avoid implying recovery of a uniquely determined 3D specimen.
  2. Important claims are insufficiently separated from reference availability, registration assumptions, and the evidence provided by the downstream tasks. The main brain benchmark uses supplied CCF pose information (Sec. 3.2), while the released count-field interface assumes that the observed section has already been registered and converted into cell-count fields; raw H&E is not itself an input to this interface. The main results therefore do not establish robustness to realistically estimated translation, rotation, or local registration errors. In addition, the downstream sparse-panel studies use separately fitted atlas-conditioned readouts. Some auxiliary experiments consume virtual 3D features, but the main downstream evidence does not by itself show that reconstructed hidden-section volumes improve the task. The more fundamental issue is the lack of published, domain-standard external baselines. The current comparisons are largely author-defined controls, including panel-only inputs, 2D versus 3D atlas variants, shuffled coordinates, central extrusion, and simple linear readouts. These controls are useful for isolating the contribution of atlas information, but they do not establish competitiveness against existing spatial-omics methods. The authors should compare with the strongest relevant published baselines for each downstream task, including Nicheformer where its input, supervision, and prediction target can be matched; where a direct comparison is infeasible, they should explain the task mismatch explicitly. Finally, the organ studies use different task-specific fitting protocols, so they support adaptability of the general principle more than cross-organ generalization of one fixed model.
Questions for authors

Q1. What unique, specimen-specific information does ATLAS S2V recover beyond a calibrated atlas? Please report gains over atlas-only at each held-out animal or section and distance, uncertainty across the four outer leave-one-animal-out folds, and residual or difference maps against matched ground truth. Q2. How would the required section-to-atlas pose be obtained for a new specimen without supplied CCF coordinates? How sensitive are the main predictions to realistic translation, rotation, and local registration errors? Q3. Which downstream experiments directly consume the predicted virtual 3D fields, and which use only observed-section or atlas-derived features? Please distinguish evidence that atlas context assists downstream inference from evidence that the reconstructed volume itself improves downstream performance. Q4. Which published methods provide the strongest relevant baselines for the hidden-gene imputation and panel cell-typing tasks? Please compare with established spatial-omics models, including Nicheformer where its task contract is compatible, using matched specimen-level splits, observed inputs, target genes or classes, and evaluation metrics. If direct comparison is not possible, please explain the mismatch in input modality, spatial resolution, supervision, or prediction target. Q5. Does the anonymous repository include the data-preparation scripts, checkpoints, and evaluation code needed to reproduce the principal results? Please identify which source datasets and intermediate caches must still be downloaded or rebuilt, and report the compute requirements for the main experiments. Q6. The released composition checkpoint appears to have its neural composition correction initialized to zero, leaving the effective mixture weight as a fixed function of absolute distance. Is the reported composition result therefore an analytic distance-gated expert mixture rather than an adaptive learned router? Please compare it with this exact analytic gate and clarify the claims about learned or specimen-conditioned routing.

4 comments

Response to Reviewer yY5R

Thank you for your questions about specimen information, registration, downstream inputs and reproducibility.

Y1. Specimen-specific information beyond a calibrated atlas

Response. The original “Atlas-only” already used the observed central section for abundance calibration; we have renamed it Calibrated atlas. AtlasS2V additionally transfers the spatial residual between the observed section and its calibrated reference. Its contribution is strongest nearby; convergence toward the calibrated atlas at greater depths is expected by design.

Distance (µm)2004001,0002,000
Calibrated atlas0.6490.6040.4890.553
AtlasS2V0.7200.6310.4890.553

Class Pearson on independent mouse C57BL6J-638850, rounded to three decimals.

New Atlas-only uses the reference alone, without central-section calibration. Four-animal leave-one-animal-out evaluation excludes each test animal from atlas construction, fitting and selection. Means equally weight animals and the shared 200/900-µm offsets.

MethodPearson ↑WAPE ↓CCC ↑NRMSE ↓
New Atlas-only0.59690.55970.53492.1280
Calibrated atlas0.62030.52050.58461.8312
AtlasS2V0.62880.50240.60651.7771

Using an atlas reference for single-section 3D prediction is itself our core innovation. The model is simple, but effective. Its contribution includes reference anatomy, specimen-specific calibration and spatial residual transfer. The gain over an already calibrated atlas measures the residual-transfer component, rather than the full benefit of introducing atlas guidance.

At shared offsets, Pearson gains over calibrated atlas are +0.0065, +0.0190, +0.0314 and -0.0230 for ABCA-1–4: 0.0085 ± 0.0233 (mean ± between-animal sample SD from rounded fold means; Appendix C.1, Table 10). Pearson improves in three animals; normalized count error decreases in all four. JSD changes from 0.0383 to 0.0399. The SD describes animal variation, not a confidence interval. Separately, removing residual transfer lowers Pearson from 0.5983 to 0.5737.

The router serves as an organ-specific adapter, as emphasized in Fig. 2. Fixed 1:1 mixing is already competitive in mouse brain. In breast, anatomical variation can limit information from a shared atlas, motivating greater reliance on the observed section. Calibrating residual strength improves density/composition Pearson from 0.8093/0.6516 to 0.8258/0.6820 over fixed 1:1 mixing. This interpretation concerns task-specific implementations across organs.

We will add matched prediction-minus-atlas, truth-minus-atlas and prediction-error maps with the same sections, classes, masks and color scales, including near/far targets and ABCA-4. These maps will show where transferred deviations agree with the measured target.

Y2. Pose estimation and registration sensitivity

Response. Across 191 sections, DeepSlice has a median 3D center error of 137.8 µm, with 94.8% within 400 µm. In separate controlled tests, AtlasS2V retains 80.8–100% of its unperturbed Pearson correlation under tested single-axis translations of 0–400 µm.

Obtaining pose without supplied coordinates. Appendix G uses the H&E image, pixel size and Allen CCFv3 template. We search coronal AP planes and in-plane rotations/reflections, refine the best candidates by constrained affine alignment of tissue outlines and internal structure, and freeze the pose before loading expression. Coronal orientation is specified from metadata or anatomical inspection; no point landmarks are supplied. Images determine pose, while molecular measurements supply the count-field adapter.

1. Registration accuracy. We evaluated all 191 sections from four datasets. Center and anterior–posterior (AP) errors are in µm; the last column gives the proportion with 3D center error ≤200 µm. Overall section-normal error was 2.88° (median) and 6.64° (P90).

DatasetSectionsCenter medianCenter P90AP median≤ 200 µm
Myelin41123.0246.373.982.9%
PITX332211.8423.0108.243.8%
Calb15896.1181.859.094.8%
Amyloid60164.1359.8120.763.3%
Overall191137.8297.177.773.8%

2. Translation sensitivity. We tested ±100–1000 µm shifts in 100 µm steps along AP, dorsal–ventral (DV) and folded medial–lateral (ML). The 0–400 µm range covers 94.8% of measured center-error magnitudes; 500–1000 µm shifts probe larger errors.

All perturbation tests use one independent mouse, the frozen model and 193 section pairs, with fixed targets and masks and no retraining or tuning. We average equally over four target distances and both perturbation signs; zero perturbation is counted once. Both sensitivity tables report AtlasS2V Pearson. Retention is Pearson divided by the unperturbed value, 0.5983.

Translation magnitude (µm)APDVFolded ML
00.59830.59830.5983
1000.58330.59160.5979
2000.54910.57200.5776
3000.51260.54210.5478
4000.48330.50870.5113
5000.34630.47650.4744
6000.32630.44660.4341
7000.30970.42030.3972
8000.29320.39840.3661
9000.32110.38000.3464
10000.30920.36410.3304

Pearson retention is 80.8–100% across tested translations up to 400 µm, and 91.8–100% up to 200 µm. At 400 µm, retention is 80.8% (AP), 85.0% (DV) and 85.5% (folded ML).

3. Orientation sensitivity. We tested ±5–45° rotations in 5° steps about AP (in-plane), DV (tilt in the original folded coordinates) and ML (tilt in the AP–DV plane). The pivot is the centroid of each observed section's valid spatial grid. Measured normal-angle errors inform the tilt tests; they do not measure in-plane rotation.

AngleIn-plane: APTilt: DVTilt: ML
0°0.59830.59830.5983
5°0.59530.58740.5727
10°0.55430.53740.5209
15°0.50700.48910.4766
20°0.46570.44510.4403
25°0.43040.40840.4120
30°0.40320.38150.3867
35°0.38290.36130.3686
40°0.36860.34380.3523
45°0.35690.33070.3377

At 5°, Pearson retention is 99.5% for in-plane rotation and 98.2%/95.7% for DV/ML tilt. Larger rotations progressively reduce performance.

4. Optional serial-section constraints. On the same 159 Myelin, Calb1 and Amyloid sections, combining angles and spacing reduces median center error from 126.1 to 111.0 µm, AP error from 75.9 to 41.3 µm and normal error from 2.88° to 1.36°. PITX3 has two cutting blocks and retains its single-section results.

Registration inputCenter medianAP medianNormal median≤ 200 µm
Independent single sections126.175.92.88°79.9%
Multi-section angle integration124.768.41.36°73.6%
Angles + spacing constraints111.041.31.36°83.0%

Positional errors are in µm; the last column refers to 3D center error. These optional constraints are not used in the 191-section single-section results.

Failure regime. At negative AP shifts of 500 µm and beyond, atlas totals become zero at some central locations, causing excessively large calibration ratios. Sign-averaged WAPE reaches approximately 3.34×10⁹–4.84×10⁹.

These tests assess global translations and rotations. Local nonrigid registration errors have not been evaluated; the Appendix G pipeline uses coronal-plane and affine alignment.

Anonymous GitHub Code: AtlasS2V-code/benchmarks/rebuttal/registration_sensitivity_0_400.zip

Y3. Atlas context versus predicted-volume context

Response. We now evaluate downstream use of the predicted 3D volume for 34-class cell typing with 10, 25 and 50 measured genes. The new readout uses context derived from the predicted volume. The table labels distinguish this evaluation from the original direct-atlas-context readout.

Method / downstream input10 genes25 genes50 genes
Genes only0.20900.47160.6366
+ 2D context0.30580.56650.6713
+ 3D atlas context (original)0.47070.66200.7094
+ Atlas and predicted-volume context (new)0.46160.65470.6997

Mean macro-F1; bold highlights the new readout.

The first two rows use measured genes alone or genes plus 2D context. The original 3D readout receives atlas context directly; the new readout additionally uses context computed from the predicted volume. Thus the table identifies the source of downstream information for each model.

The new readout improves over both gene-only and 2D inputs at every tested gene budget. At 10, 25 and 50 genes, gains over 2D context are 0.1558, 0.0882 and 0.0284, and gains over genes alone are 0.2526, 0.1831 and 0.0631. Direct 3D atlas context remains higher by 0.0091, 0.0073 and 0.0097.

Both the original and new readouts use atlas context as an additional input, which is consistent with their small performance differences. The new evaluation shows that the predicted volume can supply useful downstream context within the atlas-informed pipeline. This follows the central premise of AtlasS2V: using a biological atlas as a predictive reference when target-specimen measurements are limited, while incorporating observed-section calibration and spatial residuals.

We will distinguish the original atlas-conditioned readouts from this new volume-augmented evaluation in the text, tables and contribution statement. The largest gains over gene-only and 2D inputs occur at ten genes, where measured molecular information is most limited. This gene-budget comparison directly addresses whether the complete volume-augmented readout remains useful when few markers are measured. Its absolute macro-F1 rises from 0.4616 to 0.6997 as the gene budget increases, while the gain over 2D context decreases from 0.1558 to 0.0284.

Y4. Additional benchmark against Nicheformer and other methods

Response. We also completed a matched 10-gene, 34-class cell-typing benchmark against the Nicheformer foundation-model readout and four other published cell-typing baselines. The new AtlasS2V readout outperforms all five external baselines on both macro-F1 and accuracy.

MethodMacro-F1 ↑Accuracy ↑
Genes only0.20900.5481
Nicheformer (frozen + linear readout)0.15730.4963
CellTypist0.20770.5465
SingleR0.10070.2533
scmap-cluster0.11620.3756
TACCO (abundance adaptation)0.13480.4012
AtlasS2V: atlas + predicted-volume context0.46160.6512
Direct 3D atlas context (original control)0.47070.6832

The comparison uses four-animal leave-one-animal-out evaluation and three seeds, with the same classifier training/test cells. Scores are averaged over seeds within each animal and then equally across animals. Rejected or unassigned cells remain in the evaluation. External methods receive no additional explicit 3D context. Nicheformer uses frozen representations and a linear readout trained to convergence; this benchmark does not evaluate end-to-end fine-tuning.

AtlasS2V's macro-F1 of 0.4616 exceeds Nicheformer (0.1573) by 0.3043 and the strongest external baseline, CellTypist (0.2077), by 0.2539. The direct-atlas control remains close and slightly higher, consistent with both readouts using atlas context. These results show a substantial advantage for the complete atlas-informed pipeline under this limited-gene protocol.

Reporting both metrics separates average per-class performance from overall cell-level correctness: macro-F1 gives equal weight to each class, whereas accuracy gives equal weight to each evaluated cell. AtlasS2V also improves accuracy over Nicheformer by 0.1549 and CellTypist by 0.1047, so the advantage is present under both summaries.

The tables complement each other: Y3 compares context sources at three gene budgets, while Y4 places the complete pipeline against external methods at ten genes. The direct-atlas control remains visible in both tables to make the contribution of atlas information explicit. Nicheformer is the foundation-model baseline; CellTypist, SingleR, scmap-cluster and TACCO are additional cell-typing methods.

Anonymous GitHub Code: AtlasS2V-code/benchmarks/rebuttal/downstream_matched_04616/

Y5. Reproducibility and compute

Response. The code and frozen artifacts for the principal experiments are already available in the anonymous AtlasS2V-code repository. The relevant components can be located as follows.

  • scripts/ and src/ contain the data-preparation, atlas-construction, model-fitting, inference, and evaluation implementations. docs/experiments.md maps the principal mouse-brain, BrainCAD, organ-specific, and downstream experiments to their preparation and evaluation commands.
  • results/unified_mouse_brain_predictor_fidelity_v2/ contains the frozen composition checkpoint and composition reference. results/strict_expression_router_20260921/ contains the separately trained expression checkpoints. Artifact identities and provenance are recorded in docs/artifacts.json.
  • docs/data.md and configs/data/ specify source datasets, expected directory layouts, and specimen splits. Raw microscopy and expression datasets are obtained from their original public sources. The supplied preparation scripts generate the per-animal reference fields, native-section maps, and large expression/training caches required by the corresponding experiments.
  • requirements-tested.txt, docs/validation.json, and scripts/verify_release.py provide tested dependency versions, recorded validation results, and distributed-file checksum verification. Additional released rebuttal experiments, result tables, and reproduction instructions are organized under benchmarks/rebuttal/.

The count-field API takes registered per-class count maps. Registration and dataset-preparation scripts are in scripts/. docs/model.md specifies the composition inference rule, with numerical parity checks against the supplied PyTorch checkpoint.

Computational resources are now documented in docs/compute.md and docs/compute_downstream.md, with per-job accounting evidence and machine-readable tables under benchmarks/compute_resources/. Together, these records cover 44 workload stages and 243 successful jobs. They report available CPU allocations, runtime GPU models, wall-clock times and host-memory MaxRSS, with workload scopes identifying prepared inputs and reused artifacts. The four production stages for 25/50-gene Nicheformer embeddings and readouts sum to 12.64 wall-clock job-hours across their parallel jobs. The README and experiment guide link directly to these resource summaries.

Y6. Reference-fitted analytic distance gate

Response. The released composition predictor is an analytic distance-gated mixture with zero neural composition correction. Its fusion policy is learned from reference data during fitting and frozen at inference.

For independent-mouse evaluation, each of four reference mice is held out in turn: three build the atlas; the fourth supplies observed-center/hidden-target pairs. We compare a calibrated-atlas expert with an expert that additionally transfers the center's spatial residual. Reference targets determine the mixture weight minimizing normalized MSE at each distance; weighted least squares then fits a smooth, monotone sigmoid. Writing its intercept and positive slope as α,β\alpha,\beta,

w(d)=σ(α−β∣d∣).w(d)=\sigma(\alpha-\beta|d|).

For C57BL6J-638850, α=2.3470\alpha=2.3470 and β=8.0994 mm−1\beta=8.0994\,\mathrm{mm}^{-1}. Residual-expert weights at 200/400/1,000 µm are approximately 0.674/0.291/0.0032. The independent test mouse's hidden sections do not enter fitting.

This is reference-data-driven adaptation during fitting: expert reliability across distances determines the fusion policy. At test time, the gate is shared across positions and classes at a given distance. Specimen information enters through central-section abundance calibration and spatial residuals, changing the experts and final prediction. We will distinguish this from specimen-adaptive neural routing.

An exact analytic implementation with the same inputs and parameters is this predictor and should reproduce its outputs; released numerical parity checks compare it with the PyTorch checkpoint.

Organ/task fitting differs:

TaskWeight selection
Mouse-brain compositionReference-animal leave-one-out fitting of the distance sigmoid
BrainCAD densitySeparately trained gate with marker-specific parameters
KidneyDistance-specific fusion coefficients fitted using designated training donors
HER2ST breastReference-patient validation selects the residual model and gg; gg is constant within each outer fold, while distance enters the residual predictor

Code (under AtlasS2V-code/):

  • scripts/run_allen_langlieb_external_distance_gate.py: fit_distance_gate().
  • scripts/align_unified_composition_fidelity.py: migrate_bundle() stores fitted parameters and zeroes neural correction.
  • results/unified_mouse_brain_predictor_fidelity_v2/freeze_manifest.json: splits, per-distance weights and fitted parameters.
  • docs/model.md: inference rule.
Reviewer_twaXOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ATLASS2V, a method that predicts 3D cell type composition and cell density around a single observed tissue section. It uses a registered 3D atlas built from reference animals. The atlas is calibrated to the observed section and a residual between the two is carried to other depths. A two parameter distance gate mixes the residual expert with the atlas expert so nearby depths keep more specimen detail and far depths lean on the atlas. On one held-out mouse it beats copying the central section across offsets up to 2 mm. It also reports a four mouse cross validation and a 60 mouse density study and lung and breast and kidney studies. Downstream tests show that atlas context helps hidden gene prediction and cell typing from sparse panels.

Critical strengths
  1. I like that the problem is practical. Many spatial datasets have only one section and the idea of borrowing 3D structure from a population atlas is simple and sensible. The method has two fitted parameters and a closed form mixture so it is easy to understand and cheap to run.
  2. The paper is careful about leakage. The test animal stays out of atlas building and gate fitting and prespecified endpoints are separated from post hoc ones.
  3. The downstream cell typing result is the most convincing part. With ten genes macro F1 rises from 0.3058 to 0.4707 and all four animals improve. Shuffled depth drops performance.
Critical weaknesses
  1. Most of the gain comes from the atlas and not from the method. Atlas-only is nearly identical to ATLASS2V at far distances and a fixed 1:1 mix is within 0.003 Pearson of the full router. The headline jump from 0.392 to 0.629 is against central copy which is weak at long range.
  2. The evidence is narrow and retrospective. The main result is one mouse from the same Allen ecosystem as the atlas and the cross validation has four animals with ABCA-4 having only four predictions. Many distances are post hoc and each organ needs its own fitting.
  3. Some claims go beyond the data. The depth marginalized atlas matches the 3D query in hidden gene prediction so it does not show depth recovery. In HER2ST wrong distance performs about as well as correct distance. The claim of being first is hard to check since SpatialZ and STITCH were not run.
  4. Registration is assumed. Main benchmarks use true CCF coordinates and pose error is never injected. Real use needs image to CCF registration and robustness is unknown.
Questions for authors
  1. How much of the gain over Atlas-only survives on a second independent mouse not used in any analysis? Please report confidence intervals at the prespecified distances.
  2. Why is the gate fixed per distance instead of depending on local atlas confidence or region? Can you test region specific gates?
  3. How do results change if CCF pose is perturbed by realistic registration error such as a few hundred microns in AP or small rotations?
Reviewer_HNLBOpenReview ↗
Rating
1
Confidence
4(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AtlasS2V predicts tissue distributions at unobserved depths from a single measured section and a registered 3D atlas. It works by calibrating the atlas abundances, computing the difference between the observed section and the atlas, and then transferring that difference to other depths using a learned distance-dependent weight. The experiments have been carried out in the context of mouse brain reconstruction, marker density prediction, other organs and separate downstream models that make use of the atlas information.

Critical strengths

1. The paper has some very useful and interesting problem settings and they've done a proper and thorough evaluation.

2. The paper makes good use of publicly available atlases and datasets to evaluate their approach across a range of animals and organs.

Critical weaknesses
  • The methodological novelty is modest, just combining atlas calibration with residual weighting. The improvements are modest in these experiments: learned weighting takes Pearson from 0.5953 to 0.5983 over fixed 1:1 (Table 3). Compared to the calibrated Atlas-only component the full model gives an improvement in Pearson from 0.6203 to 0.6288 across the 4 animals and from 0.6141 to 0.6272 for the hidden layer gene expression.

  • The Method's alignment dependence as the main brain tests all assume known anatomical coordinates; the accuracy with estimated alignment is not quantitatively established.

  • The main cell-type predictor shares one decreasing distance weight across locations and cell types. This may limit its ability to capture differences in how regional or cell-type-specific deviations persist with depth.

  • I'm not sure on this one, based on my understanding the downstream benefit is unproven. The gene prediction and cell typing models just use the atlas features directly so the improvements you see in them don't demonstrate the utility of the reconstructed volume. Have you considered that? If so why have you claimed "we improve downstream inference with reconstructed 3D context"? Please clarify that part of the paper as it was unclear to me.

Questions for authors
  • Are the small gains over fixed 1:1 and calibrated Atlas-only consistent across the held out animals? How does learned weighting stack up against constant or exponential weighting that's been validated?

  • What happens to performance when the anatomical coordinates are estimated from the images? Can you quantify the sensitivity to errors in section position and orientation?

  • BrainCAD uses marker-specific weighting and tests local atlas features (Appendix C.2). Have you evaluated similar flexibility in the main cell-type reconstruction task, and does it improve performance over the shared weight?

  • Section 3.5 and Appendix F describe two separate models which use atlas features directly. How do these experiments support the assertion that "We improve downstream inference with reconstructed 3D context"? Is there any demonstrated downstream benefit from the reconstructed volume itself?

7 comments

Response to Reviewer HNLB

Thank you for highlighting the practical problem setting, use of public atlases, and evaluation across animals and organs. Your comments help us separate calibration, residual transfer, and downstream atlas conditioning.

H1a. Atlas baseline and incremental gains

Response. We first clarify an ambiguity in the original baseline name: the method previously labeled “Atlas-only” had already been calibrated using the observed central section, and therefore was not a reference-only atlas. We have renamed it Calibrated atlas to make this target-specimen information explicit.

AtlasS2V is designed to retain specimen-specific residual information near the observed section and progressively rely on the calibrated atlas at greater distances. Its convergence toward the calibrated atlas in distant sections is therefore an expected consequence of the model design. The independent-mouse results below show this pattern: the gains are larger nearby and diminish at greater depths.

Distance (µm)2004001,0002,000
Calibrated atlas0.6490.6040.4890.553
AtlasS2V0.7200.6310.4890.553

Independent mouse C57BL6J-638850; values rounded to three decimals, from the manuscript's dense-distance benchmark.

We now define Atlas-only as using the reference alone, without calibration to the observed central section. This separates the reference prior from target-informed calibration and residual transfer. In four-animal leave-one-animal-out evaluation, averaging equally over animals and the shared 200- and 900-µm offsets, the gap between this reference-only baseline and AtlasS2V is larger: mean Pearson increases from 0.5969 to 0.6288, with improvements in WAPE, CCC and NRMSE.

MethodPearson ↑WAPE ↓CCC ↑NRMSE ↓
New Atlas-only (reference only)0.59690.55970.53492.1280
Calibrated atlas0.62030.52050.58461.8312
AtlasS2V0.62880.50240.60651.7771

These comparisons clarify the roles of the individual components. Our central contribution is introducing a biological atlas as an explicit predictive reference for inferring 3D tissue distributions from a single measured section. The incremental gain over an already calibrated atlas captures one part of this contribution. Calibration and residual transfer adapt the reference to the specimen, while the router serves as an organ-specific adapter (Fig. 2) that adjusts the balance between reference guidance and observed-section information.

Finally, the breast example illustrates why this organ-specific adjustment matters. Anatomical variation between breast specimens can limit the information supplied by a shared atlas, motivating greater reliance on the observed section. In this setting, task-specific calibration of residual strength provides a larger improvement over fixed 1:1 mixing, increasing density Pearson from 0.8093 to 0.8258 and composition Pearson from 0.6516 to 0.6820. Thus, the router's central function is to adapt the balance of reference and specimen-specific information to the organ, even when equal weighting already works well in the mouse brain.

MethodDensity Pearson ↑Composition Pearson ↑
AtlasS2V0.82580.6820
Fixed g=0.5g=0.50.80930.6516
Fixed g=0.1g=0.10.78170.6019

H1b. Novelty: an atlas reference changes what is possible from one section

Response. AtlasS2V is indeed a simple model, but it is highly effective. Using an atlas reference to enable 3D tissue-distribution prediction from a single measured section is itself our core innovation. Our evaluation therefore examines the value of this reference-based formulation as a whole, beyond its incremental gains over Atlas-only.

Our contribution has three connected aspects:

  • We enable 3D tissue-distribution prediction from one target section and a registered external atlas, without additional target sections or volumetric imaging. This addresses a single-section setting left open by existing reconstruction approaches: SpatialZ generates intermediate slices between measured sections, HoloTea uses adjacent histological sections, and VORTEX requires 3D tissue morphology. AtlasS2V instead obtains the missing depth information from an external reference, extending the utility of specimens represented by only one measured section.

  • We turn shared 3D anatomy into a reusable predictive reference while preserving specimen-specific variation. The atlas supplies depth-resolved structure; the observed section calibrates cell-type abundances and identifies local spatial deviations, whose influence decreases with distance. This division of roles supports prediction beyond the measured plane while retaining individual information nearby. Across four held-out mice, mean composition Pearson increases from 0.392 to 0.629 over central-copy; residual transfer further improves the independent-mouse result from 0.5737 with the calibrated atlas to 0.5983 (component ablation).

  • We show that anatomical coverage contributes predictive information beyond the number of reference cells. Holding the reference cell budget fixed, complete coverage outperforms sparse and contiguous-gap references (mean Pearson 0.5692 versus 0.5549 and 0.5526; Appendix E.1). Within the removed interval, complete coverage raises Pearson from 0.4139 to 0.5629. This controlled ablation supports the value of preserving spatial organization when constructing a reference for single-section prediction.

We will foreground these contributions and their supporting evidence in the revised manuscript.

H2. Estimated alignment and position/orientation errors

Response. Across 191 sections, DeepSlice has a median 3D center error of 137.8 µm, with 94.8% within 400 µm. In separate controlled tests, AtlasS2V retains 80.8–100% of its unperturbed Pearson correlation under tested single-axis translations of 0–400 µm.

1. Registration accuracy. We evaluated all 191 sections from four datasets. Center and anterior–posterior (AP) errors are in µm; the last column gives the proportion with 3D center error ≤200 µm. Overall section-normal error was 2.88° (median) and 6.64° (P90).

DatasetSectionsCenter medianCenter P90AP median≤ 200 µm
Myelin41123.0246.373.982.9%
PITX332211.8423.0108.243.8%
Calb15896.1181.859.094.8%
Amyloid60164.1359.8120.763.3%
Overall191137.8297.177.773.8%

2. Translation sensitivity. We tested ±100–1000 µm shifts in 100 µm steps along AP, dorsal–ventral (DV) and folded medial–lateral (ML). The 0–400 µm range covers 94.8% of measured center-error magnitudes; 500–1000 µm shifts probe larger errors.

All perturbation tests use one independent mouse, the frozen model and 193 section pairs, with fixed targets and masks and no retraining or tuning. We average equally over four target distances and both perturbation signs; zero perturbation is counted once. Both sensitivity tables report AtlasS2V Pearson. Retention is Pearson divided by the unperturbed value, 0.5983.

Translation magnitude (µm)APDVFolded ML
00.59830.59830.5983
1000.58330.59160.5979
2000.54910.57200.5776
3000.51260.54210.5478
4000.48330.50870.5113
5000.34630.47650.4744
6000.32630.44660.4341
7000.30970.42030.3972
8000.29320.39840.3661
9000.32110.38000.3464
10000.30920.36410.3304

Pearson retention is 80.8–100% across tested translations up to 400 µm, and 91.8–100% up to 200 µm. At 400 µm, retention is 80.8% (AP), 85.0% (DV) and 85.5% (folded ML).

3. Orientation sensitivity. We tested ±5–45° rotations in 5° steps about AP (in-plane), DV (tilt in the original folded coordinates) and ML (tilt in the AP–DV plane). The pivot is the centroid of each observed section's valid spatial grid. Measured normal-angle errors inform the tilt tests; they do not measure in-plane rotation.

AngleIn-plane: APTilt: DVTilt: ML
0°0.59830.59830.5983
5°0.59530.58740.5727
10°0.55430.53740.5209
15°0.50700.48910.4766
20°0.46570.44510.4403
25°0.43040.40840.4120
30°0.40320.38150.3867
35°0.38290.36130.3686
40°0.36860.34380.3523
45°0.35690.33070.3377

At 5°, Pearson retention is 99.5% for in-plane rotation and 98.2%/95.7% for DV/ML tilt. Larger rotations progressively reduce performance.

4. Optional serial-section constraints. On the same 159 Myelin, Calb1 and Amyloid sections, combining angles and spacing reduces median center error from 126.1 to 111.0 µm, AP error from 75.9 to 41.3 µm and normal error from 2.88° to 1.36°. PITX3 has two cutting blocks and retains its single-section results.

Registration inputCenter medianAP medianNormal median≤ 200 µm
Independent single sections126.175.92.88°79.9%
Multi-section angle integration124.768.41.36°73.6%
Angles + spacing constraints111.041.31.36°83.0%

Positional errors are in µm; the last column refers to 3D center error. These optional constraints are not used in the 191-section single-section results.

Failure regime. At negative AP shifts of 500 µm and beyond, atlas totals become zero at some central locations, causing excessively large calibration ratios. Sign-averaged WAPE reaches approximately 3.34×10⁹–4.84×10⁹.

Anonymous GitHub Code: AtlasS2V-code/benchmarks/rebuttal/registration_sensitivity_0_400.zip

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

Related papers

Discussion (0)

Sign in to comment.