acceptodds
Withdrawn from ICLR 2027

ROOTFlow: 3D Tooth Root Synthesis from Intraoral Scans and Panoramic Radiographs via Latent Flow Matching

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Three-dimensional tooth surface morphology and spatial arrangement are important for dental diagnosis and treatment planning. Intraoral scans (IOS) capture detailed crown surfaces but cannot reveal tooth roots, whose three-dimensional morphology and spatial arrangement typically require cone-beam computed tomography (CBCT) to assess, at the cost of additional radiation exposure. Panoramic X-ray (PX) images provide root-related morphological and spatial cues at a lower radiation dose, but their inherent depth ambiguity and geometric distortion limit accurate three-dimensional root reconstruction. We propose ROOTFlow (Root-Oriented Oral Tomography via Latent Flow Matching), a framework that combines IOS-derived crown geometry with radiographic context to generate quadrant-level CBCT volumes and corresponding tooth segmentations, focusing on root surface morphology and spatial arrangement. The framework comprises four components: (i) quadrant-level data construction, in which unsigned distance fields (UDFs) encode IOS meshes as volumetric geometric priors; (ii) latent representation learning with two 3D variational autoencoders that compress CBCT and UDF inputs; (iii) cross-modal contrastive alignment that maps CBCT and PX features into a shared representation space; and (iv) conditional 3D latent flow matching, in which a velocity network conditioned on UDF latent features and PX features guides flow integration to generate a CBCT latent representation. The generated latent representation is jointly decoded into a quadrant-level CBCT volume and corresponding tooth segmentation. PX features provide complementary morphological and contextual cues rather than metrically reliable three-dimensional measurements. We construct a quadrant-level dataset comprising matched CBCT–IOS–PX trios derived from 5,682 patients. Evaluations using general and root-specific metrics demonstrate that ROOTFlow outperforms all evaluated baselines, with improved geometric fidelity of synthesized tooth root surfaces and more accurate tooth positioning.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D tooth root synthesis, Tooth morphology and spatial arrangement, Intraoral scans, Panoramic radiographs, Latent flow matching, Cross-modal alignment

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_JnUiOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents ROOTFlow, a framework for generating quadrant-level CBCT-like volumes and complete tooth segmentations from intraoral-scan-derived crown geometry and a same-patient panoramic radiograph. IOS crowns are represented as a truncated unsigned distance field with tooth-identity channels. Separate 3D variational autoencoders encode the IOS representation and the CBCT intensity/segmentation target. A 3D CBCT encoder and SigLIP2 panoramic encoder are trained using a symmetric contrastive objective with auxiliary tooth-presence supervision. A conditional 3D velocity network then transports the IOS latent toward the paired CBCT latent using straight-path flow matching, conditioned on IOS and panoramic global and patch-level features. The endpoint is decoded into a CBCT-like volume and per-tooth segmentation. The study uses matched CBCT, IOS, and clinical panoramic observations from 5,682 patients, with patient-level splitting and evaluation on 161 patients. ROOTFlow is compared with Oral-3D and NeBLa under real- and synthetic-panoramic configurations and with PX2Tooth for tooth geometry. It achieves the best reported whole-tooth Dice and average surface distance and outperforms the evaluated real-panoramic volumetric baselines in PSNR, SSIM, and LPIPS.

Critical strengths
  1. Large paired clinical multimodal cohort. The study uses matched CBCT, IOS, and clinical panoramic observations from 5,682 patients with patient-level data separation. Paired clinical data at this scale are difficult to obtain and allow learning from real panoramic acquisition effects rather than exclusively from CBCT-derived synthetic projections.
  2. Coherent multimodal formulation and strong reported whole-tooth performance. IOS crown geometry provides a well-motivated 3D anchor, and the complete system achieves strong whole-tooth overlap and surface-distance results. The paper distinguishes real- and synthetic-panoramic settings, reports common-tooth analyses, and provides per-position counts, improving empirical transparency.
Critical weaknesses
  1. The central tooth-root claim is not directly evaluated. The paper is framed as patient-specific tooth-root synthesis, but the primary geometric evaluations operate on whole-tooth masks. Because the patient-specific 3D crown is already observed through IOS, it may substantially determine whole-tooth Dice, ASD, PCA length, orientation, and spatial arrangement. The posterior morphological-peak descriptor is not a validated anatomical root-count measure, and the paper does not report root-only overlap or surface metrics, root-apex localization, root count, bifurcation, or curvature. Thus, the experiments demonstrate accurate IOS-anchored whole-tooth completion, but do not establish accurate recovery of the unobserved patient-specific roots. This directly affects the title, motivation, and main contribution.
  2. The experiments do not isolate panoramic-image utility or the latent-flow mechanism. There is no IOS-only baseline, correct-versus-shuffled panoramic test, alignment ablation, or direct latent-regression comparison. The model could therefore obtain strong whole-tooth results from a crown-conditioned population prior while using limited patient-specific panoramic information. Since the IOS and CBCT autoencoders are trained independently, it is also unclear why straight-line latent transport is preferable to direct conditional prediction. The main comparisons change both input information and method: ROOTFlow receives an accurate 3D IOS prior, whereas Oral-3D and NeBLa are primarily panoramic-based. They validate the complete system, but do not identify latent flow matching as the cause of the gain. The closest input setting, ToothInpaintor, is discussed but not experimentally compared.
Questions for authors

Q1 - Root-specific evaluation. Can the authors report metrics restricted to the unobserved root region, separately from the IOS-observed crown? At minimum, root-only Dice, root-surface ASD or HD95, and root-apex localization error are needed. If crown-root boundaries are not annotated, please give a reproducible separation procedure and sensitivity analysis. Root-number accuracy, molar root divergence, and root-curvature measurements would further strengthen the evaluation.

Q2 - Incremental value of the patient-specific panoramic image. Under the same architecture and training protocol, can the authors compare: (i) IOS only; (ii) IOS plus the correct panoramic image; (iii) IOS plus another patient's panorama with similar tooth presence; and (iv) IOS plus zeroed or masked panoramic features? Please report root-only metrics. This is necessary to show that the model extracts patient-specific root information rather than mainly performing tooth-identity- and crown-conditioned statistical completion.

Q3 - Necessity of alignment and flow matching. Can the authors compare no CBCT-PX contrastive alignment, global-only versus patch-only panoramic conditioning, direct CBCT-latent regression, one-step endpoint prediction, and different flow-integration step counts? Please explain why straight-line interpolation between independently learned IOS and CBCT latent spaces is appropriate. A clear advantage over direct regression is needed to substantiate the flow contribution.

Q4 - Coordinate preparation and target-derived information. For the 161 evaluation cases, how is the IOS-to-output frame obtained; does the transform use the target CBCT; do crop bounds use reference tooth labels; where does the reference image header come from without target CBCT; does panoramic cropping use manual/reference assistance; and are automatic segmentation, numbering, detection, and registration errors included? If target-derived coordinates are used, please provide an oracle-coordinate baseline and a fully target-independent evaluation.

Q5 - Symmetry of geometry evaluation. Oral-3D, NeBLa, and the reference are processed by an external tooth-instance segmenter, whereas ROOTFlow directly outputs tooth-label channels. Can the authors apply the same segmenter to ROOTFlow intensity volumes, or otherwise use an identical protocol and human-verified references for all methods? Please also report detection-aware metrics that penalize missing/extra teeth, and clarify the roles of the 1,101-patient testing set and 161-patient evaluation set, including which data informed checkpoint selection and method development.

Reviewer_vwe7OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ROOTFlow, a framework for reconstructing 3D tooth anatomy from intraoral scans (IOS) and panoramic radiographs (PX). IOS provides patient-specific crown geometry, while PX supplies complementary information about otherwise unobserved roots. The method represents crown meshes as unsigned distance fields and trains separate 3D variational autoencoders for IOS-derived inputs and CBCT targets. A contrastive alignment stage learns corresponding CBCT and radiographic representations. A conditional flow-matching network then transports the IOS latent representation toward a CBCT latent representation, conditioned on crown geometry and PX features. A decoder jointly produces a CBCT-like volume and tooth segmentations. The results support the complete reconstruction pipeline, but do not yet establish accurate recovery of isolated roots or the individual benefits of its main components.

Critical strengths
  1. Combining observed crown geometry with radiographic root information is a sensible approach to an underdetermined problem. The paper appropriately treats panoramic features as morphological evidence rather than metrically reliable 3D measurements.

  2. The cohort of 5,682 patients is a significant empirical asset. Patient-level splitting avoids the obvious leakage risk of placing quadrants from the same patient in different partitions. Including mixed dentition also broadens the anatomical variation represented during training.

  3. IOS geometry serves both as the initial flow state and as a persistent condition. Joint intensity and segmentation decoding provides complementary supervision, while the frozen decoder supplies an anatomical representation through which endpoint losses can train the velocity network.

  4. Relative to PX2Tooth, ROOTFlow improves whole-tooth DSC from approximately 0.762 to 0.835 and reduces ASD from 0.668 to 0.469 mm. Improvements also extend to length and orientation descriptors. The common-tooth analysis preserves the geometric advantage, strengthening the evidence beyond comparisons with different tooth coverage.

  5. The manuscript distinguishes real from synthetic panoramic inputs, reports valid-instance counts, and acknowledges that morphological peaks are not validated anatomical root counts. It also reports unfavorable intensity-threshold results rather than selecting only thresholds that support its ranking.

Critical weaknesses
  1. DSC and ASD are measured on whole teeth, including crowns already observed through IOS. Whole-tooth PCA axes and lengths also combine crown and root geometry. Consequently, strong scores may partly reflect accurate reconstruction of known anatomy rather than recovery of hidden roots. Root-only surface errors, apex localization, root branching, and anatomically annotated root counts are needed. The paper acknowledges this limitation, but it remains central to assessing the contribution.

  2. IOS meshes are manually registered to CBCT coordinates, training crops use tooth-label bounds, and standalone prediction reads a reference image header. These choices do not automatically invalidate the experiments, but the evaluation must specify exactly which reference-derived transforms, crop parameters, and metadata are available at prediction time. The current results do not clearly establish performance for a complete IOS–PX-only workflow.

  3. The experiments evaluate the full pipeline without isolating contrastive alignment, patient-specific PX conditioning, IOS initialization, persistent geometric conditioning, or endpoint supervision. It therefore remains unclear whether the improvement requires flow matching or could be obtained with a simpler supervised latent predictor. Unaligned PX features and shuffled-PX controls would be particularly informative.

  4. Some volumetric baselines have different input assumptions and do not exploit IOS in the same way. Their poorer performance cannot isolate the benefit of the proposed learning method from the benefit of additional geometric information. ToothInpaintor, identified as the closest input setting, is absent from the experiments. A strong baseline receiving the same IOS and PX inputs is necessary.

  5. Oral-3D, NeBLa, and reference volumes are processed by an external instance segmenter, whereas ROOTFlow also produces segmentation outputs directly. The precise source of ROOTFlow’s evaluated masks needs clarification. If its native predictions are compared with externally extracted baseline masks, the reported differences combine reconstruction quality with segmentation-pipeline effects. Agreement with independently verified annotations would strengthen the comparison.

  6. Geometry is evaluated only when a tooth is present in both prediction and reference. Missing or hallucinated teeth therefore do not directly worsen DSC or ASD. Common-tooth analysis improves comparability, but selects teeth successfully represented by every method and can exclude difficult cases. Presence accuracy alone is insufficient because it includes true negatives and accepts any nonempty prediction. Tooth-presence precision, recall, and metrics penalizing missing teeth would provide a fuller assessment.

  7. Fixed IOS and PX inputs yield a single reconstruction, despite the acknowledged possibility of multiple compatible root geometries. The method provides no calibrated indication of where the inferred anatomy is uncertain. This limits interpretation of patient-specific root predictions, especially for unusual anatomy that may be poorly represented by learned priors.

  8. The manuscript does not adequately characterize performance across scanners, institutions, ages, dentition stages, or abnormal anatomy. The role of the 1,101-patient “testing” partition versus the 161-patient “evaluation” partition also needs explanation. Confidence intervals, repeated-training variability, inference time, and memory requirements are not reported. Separate quadrant generation and heuristic overlap handling further leave full-jaw consistency unverified.

Questions for authors
  1. Can you report root-only reconstruction accuracy? In particular, please evaluate root-surface distance, apex localization, and branching or root-count accuracy using independent anatomical annotations.

  2. Which CBCT-derived information is used during evaluation preprocessing? Please distinguish training supervision from inference-time registration, crop placement, reference headers, and coordinate preparation.

  3. How much does patient-specific PX information contribute? Please compare IOS-only, PX-only, unaligned PX features, and shuffled or mismatched PX conditioning.

  4. Why is flow matching necessary? How does it compare with direct latent regression or another reconstruction model using identical inputs, representations, and supervision?

  5. Can ToothInpaintor and a strong same-input baseline be included? This would help separate architectural improvements from differences in available information.

  6. Which segmentation outputs are scored for ROOTFlow? Do its gains persist when every generated volume is processed by the same external segmenter and evaluated against human-verified reference masks?

  7. How do missing and hallucinated teeth affect overall performance? Please report presence precision and recall, subgroup coverage, and an evaluation that includes failures excluded from conditional geometry averages.

  8. How were the testing and evaluation partitions used? Please clarify model selection, provide patient-level confidence intervals, and report performance across dentition stages, acquisition settings, and difficult anatomical cases.

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

The authors propose ROOT-Flow, a framework combining IOS-driven crown geometry and Panoramic imagery which provide the complementary hidden root-related spatial cues to reconstruct Three-dimensional teeth surface and roots, obviates the need for CBCT which come at cost of high radiation exposure. Their goal is to reconstruct the quadrant level CBCT volume and their corresponding teeth segmentation. To achieve that, the paper propose a 4 stage pipeline. At start of the pipeline, it encodes the IOS Mesh into UDF volume, after that two 3D VAE encode the UDF and CBCT target quadrants into a latent space and decode them again to learn good representations, then panoramic and CBCT features are aligned together and at the last stage flow matching using velocity network conditioned on the UDF latent features and PX features, guides the flow to generate the CBCT latent features which then decoded using the CBCT VAE decoder to generate the final CBCT quadrant volume and the teeth segmentation. The framework finally evaluated on 161 cases, in terms of volume fidelity, and tooth geometry, achieving the highest Whole-tooth Dice Score, and lowest ASD among the compared methods.

Critical strengths

(1) Combining both 3D IOS Crown geometry with Panoramic images is a good and sensible multimodal approach to synthesize CBCT volumes without exposure to high radiation while acquiring the CBCT which is a need in the real world, also Decoding the CBCT latent to the CBCT quadrant and 13 teeth segmentation provides two beneficial outputs using same framework and in a single pass.

(2) The large scale of the training data, 5682 patients, ensures better generalization instead of overfitting a small dataset and the retraining of the other models (Oral-3D, NeBLa) with real and synthetic PX settings for fair comparison with the paper method .

Critical weaknesses
  1. The paper goal is to synthesize the 3D root, but it doesn't provide isolated-root evaluation, instead a whole-tooth evaluation, although the IOS crowns geometry provide high prior for the teeth information which might be the cause of the improvement of the metrics.

  2. The authors provide no ablation study at all for their framework components, the impact of the PX features, its cross-alignment with the CBCT features, Endpoint loss, and these ablation are necessary to show the effectiveness, need for each component, and whether they add value or not.

  3. The compared methods Oral-3D and NeBLa are provided with panoramic features only while RootFlow have IOS crown geometry with panoramic features which provide advantage to RootFlow over the other methods.

  4. The paper mentioned ToothInpaintor in their related work, and its input settings is the closest to them, However they didn't include it in the comparison.

  5. The Framework require registration of the IOS mesh with the CBCT target volume which shouldn't exist at inference and evaluation time.

  6. The use of external segmentation tool only on the synthesized outputs of Oral-3D and NeBLa, without applying it to the generated CBCT from RootFLow. The accuracy of the segmentation tool is not shown, and to ensure fair comparison the segmentation tool need to be used for the all generated CBCT volumes from all methods.

Questions for authors
  1. Can you provide isolated-root metrics for Dice score and ASD, instead of whole-tooth metrics by isolating the IOS crown ?

  2. Can you retrain Oral-3D and NeBLa providing the UDF volume features along with panoramic features and report again the Whole-tooth, and isolated-root metrics ?

  3. Can you perform these ablation studies: (1) IOS-only features, (2) Isolating CBCT-PX cross alignment (3) Isolating endpoint loss ?

  4. Can you make comparison between RootFLow and the other methods by applying the segmentation tool to their synthesized CBCT volumes (both RootFlow and the other methods) and provide the Whole-tooth and isolated-root Dice and ASD ?

  5. Please Clarify what you do at inference time regarding the registration of the IOS mesh.

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

Related papers

Discussion (0)

Sign in to comment.