Scaling nnU-Net for CBCT Segmentation

Constantin Ulrich; Fabian Isensee; Klaus H. Maier-Hein; Lars Kraemer; Maximilian Rokuss; Yannick Kirchhoff

REVIEW 3 major objections 4 minor 1 cited by

Scaling nnU-Net for CBCT Segmentation

T0 review · 3 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read This paper reports that a scaled nnU-Net configuration achieves state-of-the-art dental CBCT segmentation, winning the ToothFairy2 challenge with a mean Dice of 0.9253 and mean rank 4.6.

desk verdict Honest, externally backed first-place challenge report; the postprocessing cutoffs are the load-bearing weakness and need to be handled head-on. read the letter →

arxiv 2411.17213 v2 pith:ZI7SG6N5 submitted 2024-11-26 cs.CV

Fabian Isensee , Yannick Kirchhoff , Lars Kraemer , Maximilian Rokuss , Constantin Ulrich , Klaus H. Maier-Hein This is my paper · ORCID

classification cs.CV

keywords CBCTsegmentationnnU-NetToothFairy2ChallengedentalimagingdeeplearningdataaugmentationpostprocessingFDIlabeling

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

This paper reports a scaled configuration of the nnU-Net segmentation framework that wins the ToothFairy2 dental CBCT challenge. The authors enlarge the training patch to 160x320x320 voxels, deepen the network to seven resolution stages, disable left/right mirroring during training, extend training to 1500 epochs, and apply per-class postprocessing cutoffs that remove small predicted regions. On the hidden test set the method achieves a mean Dice coefficient (voxel overlap) of 0.9253 and HD95 (a 95th-percentile boundary-distance error) of 18.472, with a mean rank of 4.6 and first place among challenge teams. The paper also reports which modifications contribute most: turning off left/right mirroring gives the largest single gain, while pretraining and a separate tooth-instance segmentation stage do not beat the final pipeline once postprocessing is applied.

What carries the argument

The central mechanism is the nnU-Net ResEnc L framework—a residual-encoder U-Net with self-configuring patch size, batch size, and topology—modified in four ways: a patch of 160x320x320 voxels so the network sees enough spatial context to assign FDI tooth labels; seven resolution stages (input stride 32x64x64 in the bottleneck) instead of six; mirroring augmentation disabled only on the left/right axis because dental CBCT offers few cues for left/right disambiguation; and 1500 training epochs instead of 1000. A second load-bearing component is the postprocessing scheme: class-wise cutoffs that discard small predicted regions by turning them into background, tuned on five-fold cross-validation predictions of F cases and applied separately to Dice and HD95. The argument is that these modifications address the specific failure modes of dental CBCT—crown artifacts, fragmented teeth, and FDI misclassification—rather than requiring a new network family.

What would settle it

Run the same final pipeline with and without the postprocessing cutoffs on a held-out set of F-case CBCT scans that was never used to choose the cutoffs. If the postprocessed Dice and HD95 do not improve over the raw predictions—or if applying the cutoffs to a set with slightly different field of view or contrast reduces HD95—the reported gains would be overfitting to the training distribution rather than a generalizable improvement.

Watch

Extended reading notes

Core claim

On its own terms the contribution is a recipe, not a new architecture: the authors show that the existing nnU-Net ResEnc L model, adjusted to the geometry of dental CBCT data, wins the ToothFairy2 segmentation challenge. The central results are the measured gains from each modification—left/right mirroring off raises Dice on F-cases from 0.7445 to 0.9083 before postprocessing; adding the larger patch, longer training, and the extra resolution stage pushes postprocessed Dice to 0.9344 on F-cases, and 0.9271 on all cases in cross-validation; the final ensemble of two from-scratch models scores 0.9253 Dice and 18.472 HD95 on the test set with mean rank 4.6. The postprocessing step exploits the challenge evaluation by replacing small, likely false-positive predictions with background, with per-class cutoffs optimized separately for Dice and HD95 on F-case predictions from the five-fold cross-validation.

Load-bearing premise

The load-bearing premise is that the per-class postprocessing cutoffs, optimized on F-case predictions from the training set's five-fold cross-validation, transfer to the hidden test set, whose field of view is assumed to match the F cases; the authors note they could not independently verify this.

Editorial extensions

If this is right

The winning ToothFairy2 configuration is a ready-made baseline for future dental CBCT segmentation tasks involving 42 jaw, nerve, sinus, implant, and tooth classes.
Disabling only the left/right mirroring axis is a simple change with a large measured benefit for lateralized anatomies whose image statistics carry few left/right cues.
Metric-aware postprocessing that converts small predictions to background can improve Dice and HD95 substantially under a scoring scheme that rewards true negatives, so challenge rankings and absolute numbers should be interpreted with that scoring in mind.
Training the same model twice from scratch without fixing seeds and averaging the two outputs is a straightforward way to close the gap between single-model development scores and final test-set performance.
Because pretraining and two-stage instance segmentation did not beat the streamlined semantic pipeline after postprocessing, the main leverage for this dataset sits in configuration and postprocessing rather than in representation learning.

Reading between the lines

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

If the cutoff transfer holds beyond this challenge, the same recipe of per-class, per-metric cutoff optimization could be applied to any segmentation benchmark whose evaluation rewards true negatives, turning postprocessing into a tunable component rather than a fixed step.
The mirroring ablation suggests that other deployments on lateralized structures should consider disabling augmentation along only the left/right axis; testing this on independent dental CBCT datasets would show whether the result is specific to ToothFairy2.
The authors' explanation that reduced-field-of-view P cases interfere with tooth-instance segmentations could be tested by running that two-stage pipeline on F-only cases, where all teeth lie fully inside the image.
A nested validation, with cutoffs tuned on one fold and evaluated on another, would quantify how much of the reported cross-validation gain is optimism from tuning and evaluating on the same cases.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Bottom line: this is a clean, honest challenge report from a group that knows nnU-Net better than anyone, and the first-place claim is externally backed by the ToothFairy2 leaderboard. The genuinely new bits are small but real: disabling left/right mirroring helps for CBCT of the jaw, a 7-stage residual encoder with larger patch helps, and a class-wise postprocessing cutoff scheme tuned on cross-validation. The paper is refreshingly candid about what did not work (pretraining, full mirroring off, tooth instance segmentation) and about the optimistic nature of its postprocessed CV numbers.

Where it gets soft: the postprocessing cutoffs are fitted to the same CV predictions they are then evaluated on, and the test-set HD95 is about 3x the F-case CV value (18.47 vs 6.07). The authors admit they could not verify cutoff generalizability. Because the margin over second place is 0.2 mean rank (4.6 vs 4.8), we don't know whether the postprocessing helped, hurt, or was neutral on the hidden test set—no test-set no-PP numbers are reported. That makes the 'described pipeline as robust SOTA baseline' claim heavier than the evidence supports. The stress-test note is on point. The other soft spot is statistical: a single test set, no confidence intervals, and ranking margins that are probably within noise across reruns.

But the central result—first place—is solid as a benchmark outcome. The code is public, the cross-validation is standard 5-fold, and the external leaderboard supports the rank. This is not a methods paper proposing a new algorithm; it's a reproducible configuration report. As such, it deserves to be in the record.

Who is this for: people working on dental CBCT segmentation, and anyone using nnU-Net in structured challenges who wants to know which knobs matter. It's a useful baseline paper, not a breakthrough.

Recommendation: send it to peer review. A serious referee should ask for (a) confidence intervals or repeated runs, (b) test-set results with and without postprocessing, or at least an honest statement that the PP transfer is unverified, and (c) a small ablation isolating the PP contribution. None of that kills the paper; it makes the claims match the evidence.

Referee Report

3 major / 4 minor

Summary. This paper describes the authors' entry to the ToothFairy2 CBCT segmentation challenge. The method is based on the nnU-Net ResEnc L configuration and introduces a larger patch size (160x320x320), a deeper network (7 resolution stages), disabled left/right mirroring augmentation, longer training (1500 epochs), and class-wise postprocessing cutoffs that discard small predicted regions. On the challenge test set (n=50), the method achieves a mean Dice of 0.9253 and HD95 of 18.472, with a mean rank of 4.6, which placed first ahead of the runner-up at 4.8. The paper also reports five-fold cross-validation ablations on the training set and qualitative examples, and states that source code is publicly available.

Significance. If the reported test-set result is reliable, the contribution is a strong, reproducible baseline for multi-structure dental CBCT segmentation and a useful reference for applying nnU-Net to high-resolution 3D medical images with many classes. The paper is transparent about its design choices and failures, provides public code, and its central claim (first place in a challenge) is externally anchored by the official leaderboard. However, the paper's internal quantitative evidence is weakened by the circularity of the postprocessed cross-validation metrics and by the lack of test-set results without postprocessing, which leaves the contribution of the postprocessing component to the final ranking unverified.

major comments (3)

[Section 3.1, Table 1] The cross-validation Dice and HD95 values for the final configuration are computed on the same fold predictions that were used to optimize the class-wise postprocessing cutoffs (Section 2.2). The authors acknowledge this in the text, but Table 1 is nevertheless used to support the main ablations: for example, the improvement of the final model over the no-mirror/1500-epoch model is reported as HD95 6.069 vs 6.262 for F-cases, a margin smaller than the likely optimism in the postprocessed numbers. An unbiased estimate, obtained by optimizing cutoffs on a separate split or via nested cross-validation, is needed to support the claimed ordering of configurations.
[Section 3.4, Table 2] The test-set HD95 is 18.472, approximately three times the F-case cross-validated HD95 of 6.069, while Dice drops only modestly from 0.9344 to 0.9253. The paper attributes this drop to domain shift, but no test-set metrics without postprocessing are reported. Since the postprocessing cutoffs were optimized on F-case training predictions, the reader cannot determine whether postprocessing helped, hurt, or was neutral on the hidden test distribution. The authors note that cutoff generalizability could not be verified, but this unresolved question is load-bearing for the claim that the described pipeline, including the postprocessing component, is a robust state-of-the-art baseline. Reporting test Dice and HD95 with and without postprocessing would resolve this.
[Section 3.4] The first-place claim rests on a mean rank of 4.6 versus 4.8 for the runner-up. No confidence intervals, standard deviations, or significance tests are reported for the mean rank or for the underlying 84 per-class/metric ranks. Given the test set size (n=50) and the narrow margin, the ranking difference may be within noise. The paper should at least report the distribution of ranks or otherwise quantify the uncertainty of the mean rank.

minor comments (4)

[Section 2.2] The sentence "The cutoffs for discarding predictions are optimized on the training set predictions obtained via a five-fold cross-validation" is clear, but the subsequent sentence about choosing the smaller cutoff between Dice and HD95 is ambiguous: specify whether the cutoff is chosen per class independently for each metric and then the minimum is taken, as stated, or if some joint optimization is performed.
[Section 3.1, Table 1] The table formatting for the row with pretraining is corrupted in the manuscript ("pretr0.9119 9.2580.9320 5.875"), which makes the reported values hard to read; the table should be typeset with clear column separation.
[Section 3.3] The paper states that two models are trained from scratch on all available training data and ensembled, but it does not report whether the final test predictions come from the ensemble or a single model; this should be stated explicitly, as the cross-validation results are for single models.
[Section 2.2] The claim that test cases have the same field of view as F cases is referenced to the challenge FAQ, but the exact number of F and P cases in the training set is not given; stating this split would help the reader interpret the F-case-focused optimization.

Circularity Check

1 steps flagged · score 4.0 of 10

Postprocessed cross-validation metrics are in-sample fitted values; the test-set leaderboard claim remains independent.

fitted input called prediction [Section 2.2 'Postprocessing' and Section 3.1 'Quantitative Results (Training Set)' / Table 1]
"The cutoffs for discarding predictions are optimized on the training set predictions obtained via a five-fold cross-validation. We optimize the cutoff for each class and evaluation metric (HD95, Dice) independently. For each class, we pick the smaller cutoff between the two metrics as the final cutoff. ... We note that the reported postprocessed metrics may be optimistic because they are reported on the same cases the cutoffs were optimized on. Due to the hidden nature of the test set we were unable to independently verify the generalizability of the cutoff values."

The PP Dice/HD95 values in Table 1 are computed by applying class-wise cutoffs to the exact five-fold validation predictions on which those cutoffs were selected. Optimizing a cutoff per class and per metric and then scoring the same predictions with that metric is an in-sample fit, not an out-of-sample measurement. The reported PP improvement (e.g., F-case HD95 10.140 to 6.069) is therefore partly forced by construction. The authors explicitly concede this optimism. This circularity is confined to the internal CV ablation; the final test-set result (Dice 0.9253, HD95 18.472, mean rank 4.6) comes from a hidden leaderboard and is not circular.

full rationale

The only load-bearing circular step is the postprocessing ablation in the cross-validation table. The paper openly admits that the PP metrics are computed on the same fold predictions used to choose the per-class cutoffs, so those Dice/HD95 improvements are in-sample fitted values, not validated predictions. Everything that supports the central claim, namely first place, mean rank 4.6, and test-set Dice/HD95, comes from the hidden challenge test set and an external leaderboard, so it is independent of the fitted cutoffs. The paper's reliance on the authors' nnU-Net ResEnc L and MultiTalent prior work is not circular: nnU-Net is an established, externally validated framework, and MultiTalent pretraining is reported as a negative result. The skeptic's domain-shift observation about test HD95 being roughly three times the F-case CV value is a generalization or correctness concern, not a circularity, so it does not raise the score beyond the internal in-sample-fit issue.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new theoretical entities. It relies on two kinds of fitted quantities: postprocessing cutoffs explicitly optimized on training predictions, and manual configuration choices. The domain assumptions concern the relationship between the challenge's training and test distributions.

free parameters (2)

Class-wise postprocessing cutoffs = Not reported; one threshold per class and metric, 42 classes x 2 metrics, then min per class
Section 2.2: cutoffs are optimized on five-fold cross-validation training-set predictions. These are fitted to the training data and directly affect the reported Dice/HD95, making CV postprocessed numbers optimistic.
Manual model configuration choices = patch size 160x320x320, 7 resolution stages, 1500 epochs, left/right mirroring disabled
Patch size, depth, training duration, and mirroring settings were chosen based on past experience and CV experiments (Sections 2.2 and 3.1), not derived from first principles. The central claim depends on these choices.

assumptions (3)

domain assumption Test cases follow the same field of view as the F cases, so optimizing postprocessing cutoffs on F cases only is appropriate.
Stated in Section 2.2 and Section 3, based on challenge organizer information. If the hidden test distribution differs, the cutoffs may be miscalibrated.
domain assumption Training for 1500 epochs on 480 volumes and ensembling two unseeded models yields a hidden-test model representative of the cross-validation result.
Section 3.3. This is a standard deep learning assumption, but it is not independently validated beyond the single test set.
domain assumption nnU-Net's CT normalization and default ResEnc L topology are applicable to CBCT intensities without modification.
Section 2.1 assumes the standard nnU-Net pipeline transfers to dental CBCT. No alternative normalization or architecture adaptation is derived.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Scaling nnU-Net for CBCT Segmentation." pith.science (2026). https://pith.science/paper/ZI7SG6N5

@misc{pith2026241117213,
  author       = {Pith},
  title        = {Pith review of: Scaling nnU-Net for CBCT Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZI7SG6N5}},
  note         = {Machine review of arXiv:2411.17213}
}

read the original abstract

This paper presents our approach to scaling the nnU-Net framework for multi-structure segmentation on Cone Beam Computed Tomography (CBCT) images, specifically in the scope of the ToothFairy2 Challenge. We leveraged the nnU-Net ResEnc L model, introducing key modifications to patch size, network topology, and data augmentation strategies to address the unique challenges of dental CBCT imaging. Our method achieved a mean Dice coefficient of 0.9253 and HD95 of 18.472 on the test set, securing a mean rank of 4.6 and with it the first place in the ToothFairy2 challenge. The source code is publicly available, encouraging further research and development in the field.

Figures

Figures reproduced from arXiv: 2411.17213 by the authors.

**Figure 2.** Illustration of our method’s robust performance on challenging cases, [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Shape-aware Sampling Matters in the Modeling of Multi-Class Tubular Structures
eess.IV 2025-06 conditional novelty 5.0 of 10

A new patch-size allocation rule and skeleton-weighting scheme improves volumetric overlap and topology preservation for multi-class tubular structure segmentation in CT images.

Reference graph

Works this paper leans on

9 extracted references · 6 canonical work pages · cited by 1 Pith paper

[1]

Cipriano, S

M. Cipriano, S. Allegretti, F. Bolelli, M. Di Bartolomeo, F. Pollastri, A. Pellacani, P. Minafra, A. Anesi, and C. Grana. Deep segmentation of the mandibular canal: a new 3d annotated dataset of cbct volumes. IEEE Access , 10:11500--11510, 2022

work page 2022
[2]

Cipriano, S

M. Cipriano, S. Allegretti, F. Bolelli, F. Pollastri, and C. Grana. Improving segmentation of the inferior alveolar nerve through deep label propagation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 21137--21146, 2022

work page 2022
[3]

Isensee, P

F. Isensee, P. F. Jaeger, S. A. Kohl, J. Petersen, and K. H. Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods , 18(2):203--211, 2021

work page 2021
[4]

Isensee, T

F. Isensee, T. Wald, C. Ulrich, M. Baumgartner, S. Roy, K. Maier-Hein, and P. F. Jaeger. nnu-net revisited: A call for rigorous validation in 3d medical image segmentation. arXiv preprint arXiv:2404.09556 , 2024

arXiv 2024
[5]

Lumetti, V

L. Lumetti, V. Pipoli, F. Bolelli, E. Ficarra, and C. Grana. Enhancing patch-based learning for the segmentation of the mandibular canal. IEEE Access , 2024

work page 2024
[6]

Ronneberger, P

O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention--MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18 , pages 234--241. Springer, 2015

work page 2015
[7]

Ulrich, F

C. Ulrich, F. Isensee, T. Wald, M. Zenk, M. Baumgartner, and K. H. Maier-Hein. Multitalent: A multi-dataset approach to medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention , pages 648--658. Springer, 2023

work page 2023
[8]

, " * write output.state after.block = add.period write

ENTRY address author booktitle chapter doi edition editor eid howpublished institution journal key month note number organization pages publisher school series title type url volume year label INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1 'mid.sentence := #2 'after.sentence := #3 '...

Show all 9 references

[9]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [1]

Cipriano, S

M. Cipriano, S. Allegretti, F. Bolelli, M. Di Bartolomeo, F. Pollastri, A. Pellacani, P. Minafra, A. Anesi, and C. Grana. Deep segmentation of the mandibular canal: a new 3d annotated dataset of cbct volumes. IEEE Access , 10:11500--11510, 2022

work page 2022

[2] [2]

Cipriano, S

M. Cipriano, S. Allegretti, F. Bolelli, F. Pollastri, and C. Grana. Improving segmentation of the inferior alveolar nerve through deep label propagation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 21137--21146, 2022

work page 2022

[3] [3]

Isensee, P

F. Isensee, P. F. Jaeger, S. A. Kohl, J. Petersen, and K. H. Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods , 18(2):203--211, 2021

work page 2021

[4] [4]

Isensee, T

F. Isensee, T. Wald, C. Ulrich, M. Baumgartner, S. Roy, K. Maier-Hein, and P. F. Jaeger. nnu-net revisited: A call for rigorous validation in 3d medical image segmentation. arXiv preprint arXiv:2404.09556 , 2024

arXiv 2024

[5] [5]

Lumetti, V

L. Lumetti, V. Pipoli, F. Bolelli, E. Ficarra, and C. Grana. Enhancing patch-based learning for the segmentation of the mandibular canal. IEEE Access , 2024

work page 2024

[6] [6]

Ronneberger, P

O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention--MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18 , pages 234--241. Springer, 2015

work page 2015

[7] [7]

Ulrich, F

C. Ulrich, F. Isensee, T. Wald, M. Zenk, M. Baumgartner, and K. H. Maier-Hein. Multitalent: A multi-dataset approach to medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention , pages 648--658. Springer, 2023

work page 2023

[8] [8]

, " * write output.state after.block = add.period write

ENTRY address author booktitle chapter doi edition editor eid howpublished institution journal key month note number organization pages publisher school series title type url volume year label INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1 'mid.sentence := #2 'after.sentence := #3 '...