RAPS-3D: Efficient interactive segmentation for 3D radiological imaging

Corentin Dancette; Daniel Tordjman; Pierre Manceron; Th\'eo Danielou

REVIEW 4 major objections 7 minor 13 references

RAPS-3D: Efficient interactive segmentation for 3D radiological imaging

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that a 3D promptable segmentation model driven by a single 2D bounding box on one slice matches or exceeds prior 3D-box and text-prompt methods on the AMOS-CT benchmark while using a simpler two-pass zoom-out/zoom-in…

desk verdict A sensible, incremental simplification of SegVol whose efficiency gain is real but whose 'state-of-the-art' claim is not yet backed by matched baselines or a release. read the letter →

arxiv 2507.07730 v1 pith:YJ4TRD4O submitted 2025-07-10 cs.CV

Théo Danielou , Daniel Tordjman , Pierre Manceron , Corentin Dancette This is my paper

classification cs.CV

keywords promptablesegmentation3DmedicalimageSAMadaptationzoom-outzoom-ininferencecomputedtomographyinteractiveediting2DboundingboxpromptAMOSbenchmark

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

RAPS-3D aims to make promptable 3D segmentation practical for radiologists by replacing complicated 3D prompts and sliding-window inference with a 2D bounding box on a single slice and a two-pass zoom-out/zoom-in scheme. The paper reports mean Dice of 86.83% on AMOS-CT validation with one 2D-box prompt, higher than the 83.19% reported for SegVol's 3D box plus text and the 85.02% reported for RadSAM's 3D box, while using fewer forward passes. It also describes an interactive editing mode in which a point prompt refines the mask, reusing already computed image features when the edit point lies inside the crop. If these results hold under matched evaluation, the method would simplify interactive tools for CT organ segmentation and reduce the time cost of corrections.

What carries the argument

The load-bearing mechanism is the zoom-out/zoom-in inference scheme: the full volume is resized to the model's fixed input shape for a first pass, then the predicted region of interest is cropped, resized, and fed through the network a second time. This replaces SegVol's sliding-window pass over the cropped region and the slice-by-slice passes of autoregressive 2D methods. The prompt encoder accepts points, masks, and 2D bounding boxes, and the edition procedure reuses image features from the preceding zoom-in step when a correction point falls inside the region of interest.

What would settle it

Run both RAPS-3D and SegVol on the same AMOS-CT validation cases with identical Hounsfield normalization and prompt-format handling: if SegVol's mean Dice reaches or exceeds the reported 86.83% under that matched protocol, the state-of-the-art claim would not hold.

Watch

Extended reading notes

Core claim

The central claim is that a 3D adaptation of SAM can achieve state-of-the-art promptable organ segmentation on CT without sliding windows or autoregressive slice propagation. The model takes a volumetric image and a point or 2D bounding box, first resizes the whole volume to its input size in the zoom-out pass, then crops the region of interest indicated by the initial mask, resizes that crop, and runs a second pass in the zoom-in pass. On the AMOS-CT validation set this protocol yields a mean Dice of 86.83% with a single 2D bounding box and 88.71% after three edit points, compared with 83.19% for SegVol and 85.02% for RadSAM in the paper's comparison table.

Load-bearing premise

The headline result assumes the published baseline scores were produced under conditions equivalent to RAPS-3D's, even though SegVol's numbers come from its released demonstration pipeline, RadSAM is the same team's earlier model, and SAM-Med3D's 3-point score is missing from the comparison.

Editorial extensions

If this is right

RAPS-3D reports mean Dice 86.83% on AMOS-CT with a single 2D-box prompt, exceeding the 83.19% reported for SegVol (3D box plus text) and 85.02% for RadSAM (3D box).
Because the zoom-in step reuses features for edits inside the ROI, a correction can be computed without a second full-volume pass, cutting the latency of interactive refinement.
The two-pass inference has a fixed structure, so memory and runtime no longer scale with a sliding-window schedule; this makes deployment in clinical reading workflows simpler.
A point prompt yields 79.15% mean Dice alone and 82.98% after three edit points, showing the same architecture supports lower-cost prompts with the same editing loop.
Training on both downsampled full volumes and cropped zoom-in regions is presented as necessary for accurate initial ROIs, linking training data design to inference behavior.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the matched-protocol comparison is confirmed, the success of a single 2D box as the prompt suggests the 3D-box requirement of earlier models is not intrinsic to the task; a one-slice annotation is enough to localize a 3D structure.
The same zoom-out/zoom-in recipe may transfer to MRI and other modalities without architectural change, since the paper only evaluates CT and names MRI as future work.
A direct test of the editing design would vary the margin added when an edit point falls outside the ROI; larger margins trade speed for robustness, and the paper does not quantify that trade-off.
The efficiency gain should be volume-dependent: compared with sliding-window or per-slice methods, the fixed two-pass cost becomes more favorable as the input volume grows.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is a useful engineering paper, not a scientific breakthrough. The idea is simple—drop SegVol's sliding-window zoom-in for a single crop, and let users prompt with a 2D box on one slice—and the efficiency advantage is structural and credible. But the headline claim that this achieves state-of-the-art on AMOS-CT is not established, because the comparison table doesn't control the evaluation protocol. That doesn't sink the paper; it just means the claim needs to be reined in or backed up with a proper re-run.

What's actually good: the no-sliding-window zoom-in is a clean simplification that turns a multi-pass inference into two passes. The 2D box prompt is a real UX improvement over 3D boxes, which are awkward to specify. The editing procedure, reusing image features when the new point is inside the cropped ROI, is a sensible touch. Training on 12k public CTs and fine-tuning SegVol's ViT is a reasonable recipe, and the paper is honest about its scope—future work including more datasets and MRI is fair.

The soft spots are mostly about evidence, not method. Table 1's baseline numbers come from different prompt types, different pipelines, and at least one is from a demo. SegVol's 3D box+text result was obtained with the authors' public pipeline, not a shared evaluation script. The 3-4 point Dice margin could easily be explained by prompt protocol or preprocessing. SAM-Med3D's 3-point result is missing, leaving a gap in the bracket. There are no ablations isolating the two claimed contributions—the crop-based zoom-in and the 2D box prompt—so we don't know which part of the pipeline drives the reported gain. And there's no code or weights, so the RAPS-3D numbers can't be independently checked. These are correctable, but they're real.

The comparison to RadSAM is the same team's prior model, which is fine as a baseline, but the reader should note that the authors are comparing to their own previous work without a shared protocol either. No circular reasoning in the benchmark itself, though.

Who should read this: anyone building interactive 3D segmentation tools for clinical use. A serious referee should see it, but it needs revision: matched baselines, ablations, and a release. If the authors supply those, it becomes a solid contribution. As is, it's a promising tech report.

Referee Report

4 major / 7 minor

Summary. The paper introduces RAPS-3D, a 3D promptable segmentation model for CT volumes. It adapts a SAM-like architecture to accept a volume of shape (256,256,32) and prompts, with an emphasis on 2D bounding boxes specified on a single slice. The method uses a zoom-out/zoom-in inference scheme that explicitly avoids sliding windows, and an editing procedure that reuses previously computed features when possible. The authors fine-tune SegVol's pretrained ViT on a large collection of public CT datasets and evaluate on the AMOS-CT validation set, reporting a mean Dice of 86.83% with a single 2D bounding box, compared with 83.19% for SegVol (3D box + text) and 85.02% for RadSAM (3D box), and 88.71% after three edition points. The central claim is that this simple, efficient procedure achieves state-of-the-art performance on AMOS-CT.

Significance. If confirmed, the contribution is practically useful: a simpler inference procedure with fewer forward passes and a more intuitive 2D-box prompt would ease interactive segmentation in radiology workflows. The method is straightforward, the training data are public, and the efficiency advantage over sliding-window schemes is structural rather than hypothetical. The paper also makes a falsifiable empirical claim against external models, which is commendable. However, the significance is conditional because the state-of-the-art comparison is not protocol-matched: baseline numbers come from different prompt types, and for SegVol from a demo pipeline, while RadSAM is the same group's prior model. The reported performance advantage may partly reflect evaluation protocol rather than the method itself, so the central claim needs stronger validation before the paper's significance can be fully assessed.

major comments (4)

[Section 3, Table 1] The central state-of-the-art claim is not established because the baseline numbers are not comparable under a shared evaluation protocol. As the table note admits, the SegVol results were obtained from the authors' publicly available pipeline demonstration and released model, not from a common script. RadSAM is the same group's prior model using a 3D bounding box, SAM-Med3D's single-prompt result is for a point prompt, and its 3-point result is missing. Prompt type, HU normalization, resampling, and the set of evaluated classes can shift Dice by more than the reported 3.6-point margin (86.83 vs. 83.19). The Wilcoxon test is also reported without the number of paired samples or the unit of comparison. To support the abstract's claim, the authors should run all baselines under one common evaluation protocol (same preprocessing, same test volumes, same class scoring), or substantially qualify the claim to 'among the compared methods'.
[Section 2 (Zoom-out/Zoom-in and prompt design)] No ablation isolates the two claimed design novelties: replacing SegVol's sliding-window zoom-in with a single-crop zoom-in, and using 2D bounding boxes instead of 3D boxes or points. The reported performance could be due to the larger and differently composed training set, the fine-tuned SegVol backbone, or evaluation differences. The authors should compare RAPS-3D with and without sliding windows, and with 2D-box, 3D-box, and point prompts under identical training data, to attribute the gains to the stated components.
[Section 2.3 and Section 3] Reproducibility is insufficient: no code or model weights are provided, and several details necessary to reproduce the evaluation are missing. The paper does not specify how a 2D bounding box is defined on a single slice (e.g., which anatomical view, how the slice is chosen, how the box is converted into the 3D prompt), the exact zoom-in crop margin and resizing factors, or which AMOS-CT classes and volumes are used in the Dice calculation. Without these details and the release of code and weights, the reported state-of-the-art numbers cannot be independently verified.
[Abstract and Section 3] The 'state-of-the-art' claim is overbroad because the evaluation covers only a subset of '3D SAM's approaches' and omits strong recent methods such as nnInteractive, which also uses an auto-zoom strategy and is cited in the introduction. Including SAM-Med3D's 10-point score (83.99) does not justify a single-prompt state-of-the-art claim. The claim should be restricted to the compared 3D SAM-based methods, or additional modern baselines should be evaluated under the same protocol.

minor comments (7)

[Throughout] The text contains many missing spaces, e.g., 'Theparadigmofpromptablesegmentation' and 'T raining details', which make the paper hard to read; please fix the typography.
[Throughout] The model name is inconsistent between 'RAPS-3D' (title, abstract) and 'Raps-3D' (Section 2, Table 1, Figure 1). Use one spelling consistently.
[Section 3] The claim of a '2-3% increase in mean Dice score' from edition points is ambiguous: Table 1 shows 86.83% to 88.71%, an absolute increase of 1.88 percentage points and a relative increase of about 2.2%. Please specify which quantity is meant.
[References [8] and Table 1] RadSAM, used as a baseline, is a prior model from the same research group (reference [8] shares authors with this paper). This relationship should be disclosed in the text, and ideally the baseline should be re-run under the same evaluation protocol as RAPS-3D.
[Table 1] The table states that SAM-Med3D does not report a 3-point result, but no attempt is made to obtain one from the released model. If the authors can run the model, the missing comparison should be added; otherwise, the asymmetry between prompt types should be discussed explicitly.
[Abstract and Section 3] Dice is reported both as a fraction (0.868) and as a percentage (86.83). Standardize to one representation throughout the text and abstract to avoid ambiguity.
[Section 3 / Conclusion] Given the clinical application, the paper would benefit from a limitations paragraph mentioning out-of-distribution CTs, other modalities such as MRI, and potential failure modes; the current conclusion states this as future work but does not discuss risks.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: RAPS-3D's performance and efficiency claims are empirical and external, not derived from fitted parameters or self-cited premises.

full rationale

The paper's central claims are an empirical benchmark on AMOS-CT and a structural efficiency argument (two zoom passes instead of sliding-window inference). No equation is defined in terms of the target result, no parameter is fitted to the Dice numbers that are then reported as predictions, and the method's components are taken from external work (SegVol's pretrained ViT) rather than from a self-citation chain. The only author-overlap element, RadSAM, appears in Table 1 as a baseline obtained from a released model; that is an external empirical measurement used for comparison, not a load-bearing premise of RAPS-3D's design or derivation. Concerns about prompt mismatch or protocol comparability in Table 1 are benchmarking-validity issues, not circularity, and do not raise the circularity score. No circular steps were found.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central result rests on the trained model and the benchmark comparison. The free parameters are mostly inherited from SegVol and the training recipe, plus the qualitative crop and expand design. The key axioms are that AMOS validation is an adequate test, that the baseline numbers are comparable, and that training data do not overlap the validation set. No new scientific entities are introduced.

free parameters (5)

SegVol pretrained ViT weights (fine-tuned) = Released SegVol weights; 300-epoch fine-tune
All reported segmentation quality depends on these weights; they are not released with this paper and the fine-tuned checkpoint is not available.
Input volume size and patch size = (256,256,32) input; (16,16,4) patch
Architecture choices set without ablation; they determine how much context the zoom-out and zoom-in passes retain.
Zoom-in crop margin and resizing factors = Not specified numerically
The zoom-in ROI and edit-expansion margins are described qualitatively; exact values affect results and are absent from the text.
Training hyperparameters = LR 1e-4, batch 8, 300 epochs
Stated, but no ablations show they are optimal; they are part of the fitted system and would need to be reproduced exactly.
CT HU normalization = Clip [-500,1000], MinMax to [0,1]
Chosen normalization for CT; no comparison to alternative ranges is provided.

assumptions (4)

domain assumption Resizing and cropping full CT volumes to (256,256,32) preserves target organ structure.
Sections 2.1 and 2.3: the entire zoom-out/zoom-in pipeline assumes downsampling and crop-resize do not destroy clinically relevant anatomy.
domain assumption AMOS-CT validation is a fair external benchmark for abdominal CT promptable segmentation.
Results: the paper draws its state-of-the-art conclusion from this single validation set.
domain assumption SegVol, RadSAM, and SAM-Med3D baseline scores are comparable to RAPS-3D's runs.
Table 1: SegVol's numbers come from a demo pipeline, SAM-Med3D's 3-point result is missing, and no uniform protocol is described.
domain assumption No training-set leakage of AMOS validation cases into the 12,628 CT scans.
Section 2.3 lists public training sets but no explicit overlap check with AMOS validation is reported.

how reviews work

0 comments

Cite this review

Pith. "Pith review of RAPS-3D: Efficient interactive segmentation for 3D radiological imaging." pith.science (2026). https://pith.science/paper/YJ4TRD4O

@misc{pith2026250707730,
  author       = {Pith},
  title        = {Pith review of: RAPS-3D: Efficient interactive segmentation for 3D radiological imaging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YJ4TRD4O}},
  note         = {Machine review of arXiv:2507.07730}
}

read the original abstract

Promptable segmentation, introduced by the Segment Anything Model (SAM), is a promising approach for medical imaging, as it enables clinicians to guide and refine model predictions interactively. However, SAM's architecture is designed for 2D images and does not extend naturally to 3D volumetric data such as CT or MRI scans. Adapting 2D models to 3D typically involves autoregressive strategies, where predictions are propagated slice by slice, resulting in increased inference complexity. Processing large 3D volumes also requires significant computational resources, often leading existing 3D methods to also adopt complex strategies like sliding-window inference to manage memory usage, at the cost of longer inference times and greater implementation complexity. In this paper, we present a simplified 3D promptable segmentation method, inspired by SegVol, designed to reduce inference time and eliminate prompt management complexities associated with sliding windows while achieving state-of-the-art performance.

Figures

Figures reproduced from arXiv: 2507.07730 by the authors.

**Figure 1.** Overview of the Raps-3D inference method. (H, W, D) refers to the original image shape, and (H’, W’, D’) refers to the input model shape. arXiv:2507.07730v1 [cs.CV] 10 Jul 2025 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

13 extracted references · 4 canonical work pages

[1]

Nature communications 13(1), 4128 (2022)

Antonelli, M., Reinke, A., Bakas, S., Farahani, K., Kopp-Schneider, A., Landman, B.A., Litjens, G., Menze, B., Ronneberger, O., Summers, R.M., et al.: The medical segmentation decathlon. Nature communications 13(1), 4128 (2022)

2022
[2]

Du,Y., Bai, F., Huang, T., Zhao,B.: Segvol:Universal and interactive volumetricmedical image segmentation (2024), https://arxiv.org/abs/2311.13385

arXiv 2024
[3]

The ULS23 Challenge: a Baseline Model and Benchmark Dataset for 3D Universal Lesion Segmentation in Computed Tomography

de Grauw, M., Scholten, E.T., Smit, E., Rutten, M., Prokop, M., van Ginneken, B., Hering, A.: The uls23 challenge: a baseline model and benchmark dataset for 3d universal lesion segmentation in computed tomog- raphy. arXiv preprint arXiv:2406.05231 (2024)

work page Pith review arXiv 2024
[4]

Heller, N., Isensee, F., Trofimova, D., Tejpaul, R., Papanikolopoulos, N., Weight, C.: Kidney and Kidney Tumor Segmentation: MICCAI 2021 Challenge, KiTS 2021, Held in Conjunction with MICCAI 2021, Stras- bourg, France, September 27, 2021, Proceedings, vol. 13168. Springer Nature (2022)

work page 2022
[5]

Nature methods18(2), 203–211 (2021)

Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods18(2), 203–211 (2021)

2021
[6]

arXiv preprint arXiv:2503.08373 (2025)

Isensee, F., Rokuss, M., Krämer, L., Dinkelacker, S., Ravindran, A., Stritzke, F., Hamm, B., Wald, T., Langenberg, M., Ulrich, C., et al.: nninteractive: Redefining 3d promptable segmentation. arXiv preprint arXiv:2503.08373 (2025)

arXiv 2025
[7]

Advances in neural information processing systems35, 36722–36732 (2022)

Ji, Y., Bai, H., Ge, C., Yang, J., Zhu, Y., Zhang, R., Li, Z., Zhanng, L., Ma, W., Wan, X., et al.: Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. Advances in neural information processing systems35, 36722–36732 (2022)

work page 2022
[8]

arXiv preprint arXiv:2504.20837 (2025)

Khlaut, J., Ferreres, E., Tordjman, D., Philippe, H., Boeken, T., Manceron, P., Dancette, C.: Radsam: Segmenting 3d radiological images with a 2d promptable model. arXiv preprint arXiv:2504.20837 (2025)

arXiv 2025

Show all 13 references

[9]

Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment anything (2023), https://arxiv.org/abs/2304.02643

2023 arXiv
[10]

Medical Image Analysis97, 103285 (2024)

Li,W.,Qu,C.,Chen,X.,Bassi,P.R.,Shi,Y.,Lai,Y.,Yu,Q.,Xue,H.,Chen,Y.,Lin,X.,etal.:Abdomenatlas: A large-scale, detailed-annotated, & multi-center dataset for efficient transfer learning and open algorithmic benchmarking. Medical Image Analysis97, 103285 (2024)

2024
[11]

Wang, H., Guo, S., Ye, J., Deng, Z., Cheng, J., Li, T., Chen, J., Su, Y., Huang, Z., Shen, Y., Fu, B., Zhang, S., He, J., Qiao, Y.: Sam-med3d (2023), https://arxiv.org/abs/2310.15161

2023 arXiv
[12]

Radiology: Artificial Intelligence5(5) (Sep 2023)

Wasserthal, J., Breit, H.C., Meyer, M.T., Pradella, M., Hinck, D., Sauter, A.W., Heye, T., Boll, D.T., Cyr- iac, J., Yang, S., Bach, M., Segeroth, M.: Totalsegmentator: Robust segmentation of 104 anatomic struc- tures in ct images. Radiology: Artificial Intelligence5(5) (Sep 2...

2023 doi
[13]

arXiv preprint arXiv:2408.00874 (2024)

Zhu, J., Hamdi, A., Qi, Y., Jin, Y., Wu, J.: Medical sam 2: Segment medical images as video via segment anything model 2. arXiv preprint arXiv:2408.00874 (2024)

2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Nature communications 13(1), 4128 (2022)

Antonelli, M., Reinke, A., Bakas, S., Farahani, K., Kopp-Schneider, A., Landman, B.A., Litjens, G., Menze, B., Ronneberger, O., Summers, R.M., et al.: The medical segmentation decathlon. Nature communications 13(1), 4128 (2022)

2022

[2] [2]

Du,Y., Bai, F., Huang, T., Zhao,B.: Segvol:Universal and interactive volumetricmedical image segmentation (2024), https://arxiv.org/abs/2311.13385

arXiv 2024

[3] [3]

The ULS23 Challenge: a Baseline Model and Benchmark Dataset for 3D Universal Lesion Segmentation in Computed Tomography

de Grauw, M., Scholten, E.T., Smit, E., Rutten, M., Prokop, M., van Ginneken, B., Hering, A.: The uls23 challenge: a baseline model and benchmark dataset for 3d universal lesion segmentation in computed tomog- raphy. arXiv preprint arXiv:2406.05231 (2024)

work page Pith review arXiv 2024

[4] [4]

Heller, N., Isensee, F., Trofimova, D., Tejpaul, R., Papanikolopoulos, N., Weight, C.: Kidney and Kidney Tumor Segmentation: MICCAI 2021 Challenge, KiTS 2021, Held in Conjunction with MICCAI 2021, Stras- bourg, France, September 27, 2021, Proceedings, vol. 13168. Springer Nature (2022)

work page 2022

[5] [5]

Nature methods18(2), 203–211 (2021)

Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods18(2), 203–211 (2021)

2021

[6] [6]

arXiv preprint arXiv:2503.08373 (2025)

Isensee, F., Rokuss, M., Krämer, L., Dinkelacker, S., Ravindran, A., Stritzke, F., Hamm, B., Wald, T., Langenberg, M., Ulrich, C., et al.: nninteractive: Redefining 3d promptable segmentation. arXiv preprint arXiv:2503.08373 (2025)

arXiv 2025

[7] [7]

Advances in neural information processing systems35, 36722–36732 (2022)

Ji, Y., Bai, H., Ge, C., Yang, J., Zhu, Y., Zhang, R., Li, Z., Zhanng, L., Ma, W., Wan, X., et al.: Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. Advances in neural information processing systems35, 36722–36732 (2022)

work page 2022

[8] [8]

arXiv preprint arXiv:2504.20837 (2025)

Khlaut, J., Ferreres, E., Tordjman, D., Philippe, H., Boeken, T., Manceron, P., Dancette, C.: Radsam: Segmenting 3d radiological images with a 2d promptable model. arXiv preprint arXiv:2504.20837 (2025)

arXiv 2025