Writing study: A Scalable Distributed Framework for Multimodal GigaVoxel Image Registration

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Rohit Jena, Vedant Zope, Pratik A Chaudhari, James Gee · Official ICLR 2026 oral record

Observed writing choices, not acceptance rules. This study examines the inspected manuscript version. It does not establish why the paper was accepted or that its presentation choices should be copied unchanged.

Figure, table and section ledger

Editorial contribution type: Inference, algorithms & systems efficiency (primary).

Layout extraction, not a required paper template. Counts are distinct caption labels; a multi-panel figure counts once. Section spans are estimated from heading positions, including floats and whitespace. Method and limitations.

Canonical PDF

20 physical pages

Before references / appendix

≈ 10.18 page-equivalents

Main figures / tables

8 / 1

Attached appendix figures / tables

0 / 0

Attached appendix span

Separate file recorded below; not in this PDF total

Before references / appendix: ≈ 50.9%References / intervening material: ≈ 49.1%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Extraction notes: No attached appendix boundary detected. Zero means no recognized caption, not independently verified absence.

Source-bound correction: Source main headings use Title Case rather than small-cap uppercase. All six numbered top-level headings manually identified; locations are source-pinned Poppler bbox coordinates. No integrated appendix; separate official supplement holds appendices.

Section-by-section lengths and source pages

The title/abstract/front matter precede the first numbered heading. A span ends at the next detected top-level heading, not at the last sentence; end-page whitespace is included. References and statements are listed separately. These are not measurements of text-only or artwork area.

Top-level heading	Starts on PDF page	Estimated span	Share of entire PDF
1 Introduction	1	≈ 1.55 pp.	7.75%
2 Related work	3	≈ 0.65 pp.	3.25%
3 Fused kernels for memory efficient registration on a single GPU	3	≈ 2.12 pp.	10.6%
4 Extending image registration to multiple GPUs	5	≈ 2.4 pp.	12%
5 Experiments	8	≈ 2.46 pp.	12.3%
6 Conclusion	10	≈ 0.38 pp.	1.9%
Acknowledgments	11	≈ 0.07 pp.	0.35%
References	11	≈ 9.82 pp.	49.1%

All detected figure and table locations

Main text

• Figure 1 · p. 2
• Figure 2 · p. 4
• Figure 3 · p. 5
• Figure 4 · p. 6
• Figure 5 · p. 7
• Figure 6 · p. 8
• Figure 7 · p. 9
• Table 1 · p. 9
• Figure 8 · p. 10

Attached appendix

No recognized caption labels.

Separate source files—excluded from canonical-PDF totals

A separate author version may repeat main text and references. Whole-file lengths and counts are not extra appendix material and are never added to the canonical PDF totals. Read the study's source scope for which sections received substantive reading versus layout inspection.

Additional source 1

Separate PDF (or ZIP containing it): whole-file layout of 26 physical pages; 12 detected figures and 4 tables. These may include overlapping main-text material.

Located appendix only: starts on physical p.1; ≈ 25.89 page-equivalents, 12 figure labels and 4 table labels. This interval is distinct from the whole file.

Detected heading	PDF page	Estimated span
A Appendix	1	≈ 0.03 pp.
B Related works	1	≈ 3.36 pp.
C Limitations and future work	4	≈ 0.22 pp.
D LLM usage	4	≈ 0.08 pp.
E Correctness of grid parallel implementation	4	≈ 1.98 pp.
F Accelerating TransMorph training	6	≈ 1.57 pp.
G An efficient fused local normalized cross correlation loss	8	≈ 3.76 pp.
H A highly efficient mutual information implementation	12	≈ 1.72 pp.
I Composite implicit grid sampler	13	≈ 0.03 pp.
J Ring sampler for scalable distributed interpolation	13	≈ 3.31 pp.
K Correctness of implementation	17	≈ 1.44 pp.
L Additional details on the simulated ex-vivo brain MRI dataset	18	≈ 8.39 pp.

• Figure 9 · p. 5
• Figure 10 · p. 6
• Figure 11 · p. 7
• Figure 12 · p. 7
• Table 2 · p. 8
• Figure 13 · p. 9
• Table 3 · p. 11
• Figure 14 · p. 16
• Table 4 · p. 17
• Table 5 · p. 18
• Figure 15 · p. 19
• Figure 16 · p. 23
• Figure 17 · p. 24
• Figure 18 · p. 25
• Figure 19 · p. 26
• Figure 20 · p. 26

Entire source is a separate 26-page appendix with Title Case top-level headings. Printed page21 is physical page1. Figure9/Table2 continue the main-paper numbering, so missing lower labels are not extraction failures.

Argument and organization

1. Bottlenecks before kernels

   Uses the registration objective and memory profiling to justify three fused operations.

2. Single-device to distributed design

   Introduces grid sharding and interpolation communication only after reducing local memory costs.

3. Feasibility, accuracy and implementation

   Separates real native-resolution registration, synthetic accuracy tests and reproducible kernel derivations.

Problem, claim and contribution

The introduction translates acquisition resolution into transformation-parameter and memory requirements, making the engineering gap tangible without introducing an unnecessary new learning model.

Sections 1 and 3, pages 1–3 · Main paper · Additional source 2

Argument and section order

The progression from fused single-GPU primitives to GridParallel and Ring Sampler mirrors a reader's questions: reduce intermediates first, then resolve boundary synchronization and remote sampling.

Sections 3–4, pages 3–7 · Main paper · Additional source 2

Figures, captions and visual explanation

Figure 2 gives a color-coded memory hierarchy and a before/after computation diagram alongside bottleneck shares. It links the implementation change to the resource that is actually scarce.

Figure 2, page 4 · Main paper · Additional source 2

Tables, comparisons and evidence

Figure 5 groups rows by physical resolution and marks patch-based adaptations. Adjacent accuracy–cost plots expose tradeoffs that would disappear in a single speedup number.

Figure 5, page 7 · Main paper · Additional source 2

Appendix and reproducibility

The supplement provides forward/backward pseudocode, dataset synthesis, method-specific settings, metric definitions and a weak-scaling limitation. It explicitly discloses the approximate ANTs gradient option used in experiments.

Supplement Appendices C, G, J and L, printed pages 24, 28–36 and 38–42 · Main paper · Additional source 2

Clarity and writing conventions

Several main-text appendix references are shifted relative to the separate supplement's lettering. Terms such as arbitrary scale and exactness need qualification when the implementation uses approximations and finite communication bandwidth.

Main Sections 3–5; supplement Appendices G.1 and H.1.2 · Main paper · Additional source 2

What to reuse—and what not to infer

Transferable lesson: Build a systems paper around a measured bottleneck, show the memory movement visually, and couple each scale claim to accuracy, hardware, approximation flags and communication cost.

Boundary: MRI and mouse-brain microscopy results do not establish WSI-registration accuracy. Faux-OASIS is synthesized from coarse labels, and weak scaling is only 41% on eight GPUs without NVLink/InfiniBand.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-09-19

Reading status

Full text + selected visuals inspected

Main text

Pages 1–10, Sections 1–6; bibliography occupies pages 11–20

Appendix / supplement

Complete separate supplement, PDF pages 1–26 / printed pages 21–46, Appendices A–L

Visually inspected PDF pages

4, 7

• ICLR 2026 final proceedings · 20 pages
• ICLR 2026 final supplement · 26 pages