arxiv: 2604.20213 · v1 · submitted 2026-04-22 · 💻 cs.CV

Recognition: unknown

Weighted Knowledge Distillation for Semi-Supervised Segmentation of Maxillary Sinus in Panoramic X-ray Images

Juha Park , Jiho Choi , Jong Pil Yun , Yong Chan Park , Han-Gyeol Yeom , Byung Do Lee , Sang Jun Lee

Authors on Pith no claims yet

Pith reviewed 2026-05-10 00:52 UTC · model grok-4.3

classification 💻 cs.CV

keywords semi-supervised segmentationmaxillary sinuspanoramic X-rayknowledge distillationCycle-GANdental imagingpseudo-label refinementimage segmentation

0 comments

The pith

A weighted knowledge distillation loss paired with SinusCycle-GAN refinement delivers 96.35% Dice score for maxillary sinus segmentation under limited labels.

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

The paper presents a semi-supervised framework to segment the maxillary sinus in panoramic X-ray images when few expert-labeled scans are available. A teacher model generates pseudo labels that a student learns from, but a weighted distillation loss reduces the influence of regions where the two models disagree structurally. A separate SinusCycle-GAN then refines those pseudo labels by translating images without paired examples, sharpening boundaries and lowering noise. The resulting model is shown to exceed prior segmentation methods on a clinical collection of 2,511 patients while producing more consistent anatomical outlines. This matters for dental diagnosis and planning because accurate sinus maps can be obtained without requiring exhaustive manual annotation of every new image.

Core claim

By introducing a weighted knowledge distillation loss that suppresses unreliable signals arising from teacher-student structural discrepancies and by applying a SinusCycle-GAN refinement network based on unpaired image-to-image translation to improve pseudo-label boundaries and reduce noise, the semi-supervised framework achieves a Dice score of 96.35% on maxillary sinus segmentation in panoramic radiographs, outperforming state-of-the-art models and lowering boundary error even when labeled data are scarce.

What carries the argument

Weighted knowledge distillation loss that down-weights mismatched teacher-student regions, combined with SinusCycle-GAN for refining pseudo labels via unpaired translation.

If this is right

The method yields anatomically consistent segmentations suitable for dental diagnosis and surgical planning even with limited annotations.
Boundary error is reduced relative to existing segmentation models on the same panoramic X-ray collection.
Noise propagation from unlabeled data is limited during training.
The framework supports broader dental image analysis tasks under similar data constraints.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same weighting and refinement strategy could be tested on other 2D projections with overlapping anatomy, such as chest radiographs.
Pairing the approach with 3D CT volumes might further reduce errors in complex sinus cases.
The unpaired refinement step could be adapted to other medical modalities that suffer from domain shift between labeled and unlabeled sets.
Performance across different patient demographics or scanner types would need verification to confirm general robustness.

Load-bearing premise

The weighted distillation loss and SinusCycle-GAN reliably improve pseudo-label quality and suppress unreliable signals without introducing new biases or artifacts.

What would settle it

On the collected 2,511-patient dataset, if expert-annotated test masks show that the refined pseudo labels have higher boundary disagreement than the original teacher outputs, or if the full method's Dice score falls below a standard supervised baseline trained on the same labeled subset.

Figures

Figures reproduced from arXiv: 2604.20213 by Byung Do Lee, Han-Gyeol Yeom, Jiho Choi, Jong Pil Yun, Juha Park, Sang Jun Lee, Yong Chan Park.

**Figure 1.** Figure 1: Overview of the proposed method. Panoramic X-ray images are fed into both the teacher and student models. The teacher model is trained using only labeled data, whereas the student model is trained using both labeled and unlabeled data. After the teacher model is trained, it generates pseudo labels for the unlabeled data. These pseudo labels are refined by SinusCycle-GAN and subsequently used to train the s… view at source ↗

**Figure 2.** Figure 2: Architecture of the proposed SinusCycle-GAN for pseudo label refinement. The network is based on an unpaired image-to-image translation framework and incorporates correction network with an encoder-decoder architecture with CBAM to enhance boundary accuracy and suppress noise in initial pseudo labels ˆyt . The refined output is ˜yt and corresponds to the high-quality pseudo label used to train the student … view at source ↗

**Figure 3.** Figure 3: Examples of panoramic X-ray images for maxillary sinus segmentation. (a) panoramic X-ray images. (b) ground truth masks annotated by expert [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: Number of subjects by age group and sex in the maxillary sinus clinical dataset. surgery and with either normal maxillary sinus mucosa or only mild mucosal thickening observed on panoramic radiographs. Patients presenting with congenital jaw deformities or documented maxillofacial trauma that could compromise reliable radiographic interpretation were excluded. As illustrated in [PITH_FULL_IMAGE:figures/fu… view at source ↗

**Figure 5.** Figure 5: Qualitative comparison of maxillary sinus segmentation results obtained using different methods. From left to right, the columns show the input panoramic X-ray image, TransUNet, UNETR, nnU-Net, MedSAM, the proposed method, and the ground truth mask. The proposed method produces smoother and more anatomically consistent boundaries compared with the other methods. local discontinuities remain evident when ap… view at source ↗

**Figure 6.** Figure 6: Representative failure cases of the proposed method for maxillary sinus segmentation in panoramic X-ray images. (a) internal holes. (b) underestimation of the ground truth region. (c) overestimation of the ground truth region. and overlapping anatomical structures, where local boundary cues are insufficient for precise delineation. Nevertheless, such failure cases are observed in a relatively small number … view at source ↗

read the original abstract

Accurate segmentation of maxillary sinus in panoramic X-ray images is essential for dental diagnosis and surgical planning; however, this task remains relatively underexplored in dental imaging research. Structural overlap, ambiguous anatomical boundaries inherent to two-dimensional panoramic projections, and the limited availability of large scale clinical datasets with reliable pixel-level annotations make the development and evaluation of segmentation models challenging. To address these challenges, we propose a semi-supervised segmentation framework that effectively leverages both labeled and unlabeled panoramic radiographs, where knowledge distillation is utilized to train a student model with reliable structural information distilled from a teacher model. Specifically, we introduce a weighted knowledge distillation loss to suppress unreliable distillation signals caused by structural discrepancies between teacher and student predictions. To further enhance the quality of pseudo labels generated by the teacher network, we introduce SinusCycle-GAN which is a refinement network based on unpaired image-to-image translation. This refinement process improves the precision of boundaries and reduces noise propagation when learning from unlabeled data during semi-supervised training. To evaluate the proposed method, we collected clinical panoramic X-ray images from 2,511 patients, and experimental results demonstrate that the proposed method outperforms state-of-the-art segmentation models, achieving the Dice score of 96.35\% while reducing boundary error. The results indicate that the proposed semi-supervised framework provides robust and anatomically consistent segmentation performance under limited labeled data conditions, highlighting its potential for broader dental image analysis applications.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This paper offers a practical semi-supervised method for dental X-ray segmentation with some novel tweaks to distillation and GAN-based refinement, but the results need stronger validation to confirm the proposed parts are driving the improvements.

read the letter

The main point for you is that this work adapts knowledge distillation with a weighting mechanism and introduces a SinusCycle-GAN to handle pseudo-label refinement in a semi-supervised setup for maxillary sinus segmentation on panoramic X-rays, reporting a high Dice score of 96.35 percent on their collected dataset of over 2500 patients. They do something useful by focusing on a niche but clinically relevant task where structural overlaps in 2D images make segmentation tricky and labeled data is scarce. The weighted loss aims to reduce the impact of discrepancies between teacher and student models, and the GAN is meant to clean up the pseudo labels for better boundary accuracy. Collecting real clinical images from 2511 patients is a solid step, and showing outperformance over existing models under semi-supervised conditions adds practical value. That said, the evidence for the new components working as intended is not fully convincing yet. The description lacks specifics on how the weighting is calculated, what the exact architecture of the SinusCycle-GAN is, or any ablation studies isolating their contributions. More importantly, there's no mention of validating the refined pseudo labels against expert annotations or analyzing cases where the methods might introduce artifacts due to the ambiguous boundaries common in these X-rays. Without statistical tests or failure case discussions, the boundary error reduction could be overstated or influenced by other aspects of the training. This kind of paper is for specialists in medical imaging, particularly those dealing with dental radiology or semi-supervised learning in constrained data settings. Someone building on similar distillation techniques might find the adaptations worth looking at, but it won't change the broader field. It deserves to go through peer review because the dataset size and the applied focus make it worth a closer look, even if the authors will need to add more rigorous validation to back up the claims.

Referee Report

3 major / 1 minor

Summary. The paper claims to introduce a semi-supervised segmentation framework for maxillary sinus in panoramic X-ray images. It employs a teacher-student knowledge distillation setup with a weighted loss to suppress unreliable signals arising from structural discrepancies between predictions, and introduces SinusCycle-GAN, an unpaired image-to-image translation network, to refine pseudo-labels generated from unlabeled data. On a private clinical dataset of panoramic radiographs from 2,511 patients, the method is reported to outperform state-of-the-art segmentation models, achieving a Dice score of 96.35% with reduced boundary error, and to provide robust performance under limited labeled data conditions.

Significance. If the results hold after proper validation, the work could be significant for semi-supervised medical image segmentation in dental radiography, where annotation costs are high and 2D projections create ambiguous boundaries. The domain-specific adaptations (weighted distillation and SinusCycle-GAN refinement) target practical challenges in pseudo-label quality, potentially reducing reliance on large labeled sets while maintaining anatomical consistency. The scale of the patient cohort is a positive factor, though impact depends on reproducibility and checks against artifacts.

major comments (3)

[Abstract] Abstract: The headline performance claim (96.35% Dice, reduced boundary error, SOTA outperformance) is presented without any description of the experimental protocol, including the labeled/unlabeled split, choice of baselines, evaluation metrics for boundary error, or statistical significance testing. This information is load-bearing for assessing whether the weighted KD and SinusCycle-GAN mechanisms deliver the reported gains or merely correlate with them.
[Methods] Methods (SinusCycle-GAN description): The claim that SinusCycle-GAN produces anatomically faithful pseudo-labels and reduces noise propagation rests on the assumption that unpaired image-to-image translation preserves maxillary sinus contours in the presence of structural overlap. No quantitative checks (e.g., expert review of refined labels on a held-out subset or failure-case analysis) are referenced, leaving open the risk that the module introduces boundary artifacts or biases as noted in the stress-test.
[Results] Results: The central assertion that the weighted knowledge distillation loss reliably suppresses unreliable teacher-student discrepancies lacks supporting ablation studies or analysis of how the weighting affects pseudo-label quality. Without these, it is difficult to confirm that the mechanism improves rather than masks errors in ambiguous panoramic projections.

minor comments (1)

[Abstract] Abstract: The phrase 'reducing boundary error' is used without naming the specific metric (e.g., Hausdorff distance, average surface distance), which would aid clarity and allow direct comparison to prior work.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below, offering clarifications from the manuscript and proposing targeted revisions where they will strengthen the presentation without altering the core contributions.

read point-by-point responses

Referee: [Abstract] Abstract: The headline performance claim (96.35% Dice, reduced boundary error, SOTA outperformance) is presented without any description of the experimental protocol, including the labeled/unlabeled split, choice of baselines, evaluation metrics for boundary error, or statistical significance testing. This information is load-bearing for assessing whether the weighted KD and SinusCycle-GAN mechanisms deliver the reported gains or merely correlate with them.

Authors: We agree that the abstract would be more informative with a concise reference to the experimental protocol. The full manuscript (Section 4) details the semi-supervised split on the 2,511-patient cohort, the specific SOTA baselines, boundary error metrics (including Hausdorff distance), and statistical testing. In revision we will expand the abstract by one sentence summarizing these elements to better contextualize the 96.35% Dice result and the role of the proposed components. revision: yes
Referee: [Methods] Methods (SinusCycle-GAN description): The claim that SinusCycle-GAN produces anatomically faithful pseudo-labels and reduces noise propagation rests on the assumption that unpaired image-to-image translation preserves maxillary sinus contours in the presence of structural overlap. No quantitative checks (e.g., expert review of refined labels on a held-out subset or failure-case analysis) are referenced, leaving open the risk that the module introduces boundary artifacts or biases as noted in the stress-test.

Authors: The manuscript supports the claim through qualitative refinement examples and the stress-test analysis that evaluates robustness to structural overlap. We acknowledge that a dedicated quantitative expert review on refined labels is not presented. In revision we will add an explicit paragraph in the Methods or Discussion section discussing the assumptions of unpaired translation, referencing the existing stress-test results, and noting potential boundary artifacts as a limitation. revision: partial
Referee: [Results] Results: The central assertion that the weighted knowledge distillation loss reliably suppresses unreliable teacher-student discrepancies lacks supporting ablation studies or analysis of how the weighting affects pseudo-label quality. Without these, it is difficult to confirm that the mechanism improves rather than masks errors in ambiguous panoramic projections.

Authors: The results section already includes comparative experiments contrasting the weighted KD loss against unweighted variants, showing consistent gains in Dice and boundary metrics, together with visual pseudo-label comparisons. To directly address the request for more granular analysis, we will expand the ablation subsection with additional quantitative breakdowns of weighting effects on pseudo-label quality. revision: yes

Circularity Check

0 steps flagged

Empirical framework with no derivations or self-referential reductions

full rationale

The paper is entirely empirical: it proposes a semi-supervised segmentation pipeline (weighted KD loss + SinusCycle-GAN refinement) and reports experimental Dice scores on a 2,511-patient dataset. No equations, first-principles derivations, parameter-fitting steps presented as predictions, or uniqueness theorems appear in the abstract or described claims. The central results are performance numbers obtained from training and evaluation; they do not reduce by construction to the inputs via self-definition, fitted-input renaming, or self-citation chains. Any self-citations that may exist in the full text would be non-load-bearing for a derivation that does not exist.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

The central claim rests on the effectiveness of the proposed weighted knowledge distillation and the SinusCycle-GAN, which are presented as novel but their internal hyperparameters and training details are not specified in the abstract. Standard assumptions in deep learning for segmentation apply.

invented entities (1)

SinusCycle-GAN no independent evidence
purpose: refinement network based on unpaired image-to-image translation to improve pseudo labels
Introduced as a new component in the semi-supervised framework to enhance boundary precision and reduce noise.

pith-pipeline@v0.9.0 · 5576 in / 1309 out tokens · 67454 ms · 2026-05-10T00:52:37.846903+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

30 extracted references · 21 canonical work pages · 2 internal anchors

[1]

M., Lu, H

Sala, Y . M., Lu, H. & Chrcanovic, B. R. Clinical outcomes of maxillary sinus floor perforation by dental implants and sinus membrane perforation during sinus augmentation: a systematic review and meta-analysis.J. Clin. Medicine13, 1253, DOI: 10.3390/jcm13051253 (2024). 2.Rushton, V . E. & Horner, K. The use of panoramic radiology in dental practice.J. De...

work page doi:10.3390/jcm13051253 2024
[2]

Malina-Altzinger, J., Damerau, G., Grätz, K. W. & Stadlinger, P. B. Evaluation of the maxillary sinus in panoramic radiography—a comparative study.Int. journal implant dentistry1, 17 (2015)

2015
[3]

Reports12, 7523, DOI: 10.1038/s41598-022-11589-9 (2022)

Morgan, N.et al.Convolutional neural network for automatic maxillary sinus segmentation on cone-beam computed tomographic images.Sci. Reports12, 7523, DOI: 10.1038/s41598-022-11589-9 (2022)

work page doi:10.1038/s41598-022-11589-9 2022
[4]

In: Medical Image Compu ting and Computer-Assisted Intervention – MICCAI 2015

Ronneberger, O., Fischer, P. & Brox, T. U-net: Convolutional networks for biomedical image segmentation.Int. Conf. on Med. Image Comput. Comput. Interv. (MICCAI)234–241, DOI: 10.1007/978-3-319-24574-4_28 (2015)

work page doi:10.1007/978-3-319-24574-4_28 2015
[5]

L.et al.Accurate segmentation of dental panoramic radiographs with u-nets.2019 IEEE 16th Int

Koch, T. L.et al.Accurate segmentation of dental panoramic radiographs with u-nets.2019 IEEE 16th Int. Symp. on Biomed. Imaging (ISBI 2019)1–4 (2019)

2019
[6]

Sci.13, 1516, DOI: 10.3390/app13031516 (2023)

Ba-Hattab, R.et al.Detection of periapical lesions on panoramic radiographs using deep learning.Appl. Sci.13, 1516, DOI: 10.3390/app13031516 (2023)

work page doi:10.3390/app13031516 2023
[7]

Hiraiwa, T.et al.A deep-learning artificial intelligence system for assessment of root morphology of the mandibular first molar on panoramic radiography.Dentomaxillofacial Radiol.48, 20180218, DOI: 10.1259/dmfr.20180218 (2019)

work page doi:10.1259/dmfr.20180218 2019
[8]

Reports9, 8495, DOI: 10.1038/s41598-019-44839-0 (2019)

Krois, J.et al.Deep learning for the radiographic detection of periodontal bone loss.Sci. Reports9, 8495, DOI: 10.1038/s41598-019-44839-0 (2019)

work page doi:10.1038/s41598-019-44839-0 2019
[9]

Bonfanti-Gris, M.et al.Deep learning for tooth detection and segmentation in panoramic radiographs: A systematic review and meta-analysis.BMC Oral Heal.25, 1280, DOI: 10.1186/s12903-025-03864-2 (2025)

work page doi:10.1186/s12903-025-03864-2 2025
[10]

& Suebnukarn, S

Poedjiastoeti, W. & Suebnukarn, S. Application of convolutional neural network in the diagnosis of jaw tumors.Healthc. Informatics Res.24, 236–241, DOI: 10.4258/hir.2018.24.3.236 (2018)

work page doi:10.4258/hir.2018.24.3.236 2018
[11]

Sci.11, 11904, DOI: 10.3390/app112411904 (2021)

Huang, Y .-C.et al.Tooth position determination by automatic cutting and marking of dental panoramic x-ray film in medical image processing.Appl. Sci.11, 11904, DOI: 10.3390/app112411904 (2021)

work page doi:10.3390/app112411904 2021
[12]

& Viriri, S

Majanga, V . & Viriri, S. Dental images’ segmentation using threshold connected component analysis.Comput. Intell. Neurosci.2021, 2921508, DOI: 10.1155/2021/2921508 (2021)

work page doi:10.1155/2021/2921508 2021
[13]

Cha, J.-Y .et al.Panoptic segmentation on panoramic radiographs: Deep learning-based segmentation of various structures including maxillary sinus and mandibular canal.J. Clin. Medicine10, 2577, DOI: 10.3390/jcm10122577 (2021)

work page doi:10.3390/jcm10122577 2021
[14]

Kwon, O.et al.Automatic diagnosis for cysts and tumors of both jaws on panoramic radiographs using a deep convolution neural network.Dentomaxillofacial Radiol.49, 20200185, DOI: 10.1259/dmfr.20200185 (2020)

work page doi:10.1259/dmfr.20200185 2020
[15]

Oral Radiol.129, 635–642, DOI: 10.1016/j.oooo.2019

Lee, J.-H.et al.Application of a fully deep convolutional neural network to the automation of tooth segmentation on panoramic radiographs.Oral Surgery, Oral Medicine, Oral Pathol. Oral Radiol.129, 635–642, DOI: 10.1016/j.oooo.2019. 11.007 (2020)

work page doi:10.1016/j.oooo.2019 2019
[16]

Murata, M.et al.Deep-learning classification using convolutional neural network for evaluation of maxillary sinusitis on panoramic radiography.Oral Radiol.35, 301–307, DOI: 10.1007/s11282-019-00371-0 (2019)

work page doi:10.1007/s11282-019-00371-0 2019
[17]

Distilling the Knowledge in a Neural Network

Hinton, G., Vinyals, O. & Dean, J. Distilling the knowledge in a neural network.arXiv preprint arXiv:1503.02531(2015)

work page internal anchor Pith review Pith/arXiv arXiv 2015
[18]

& Luo, J

Zheng, H., Fu, J., Mei, T. & Luo, J. Looking for the devil in the details: Learning trilinear attention sampling network for fine-grained image recognition.Proc. IEEE/CVF Conf. on Comput. Vis. Pattern Recognit. (CVPR)5012–5021, DOI: 10.1109/CVPR.2019.00515 (2019)

work page doi:10.1109/cvpr.2019.00515 2019
[19]

MViTv2: Improved Multiscale Vision Transformers for Classification and Detection , isbn =

Yang, C.et al.Cross-image relational knowledge distillation for semantic segmentation.Proc. IEEE/CVF Conf. on Comput. Vis. Pattern Recognit. (CVPR)12319–12328, DOI: 10.1109/CVPR52688.2022.01203 (2022)

work page doi:10.1109/cvpr52688.2022.01203 2022
[20]

Dn-splatter: Depth and normal priors for gaussian splatting and meshing

Liu, L.et al.Bpkd: Boundary privileged knowledge distillation for semantic segmentation.Proc. IEEE/CVF Winter Conf. on Appl. Comput. Vis. (WACV)DOI: 10.1109/W ACV57701.2024.00462 (2024)

work page doi:10.1109/w 2024
[21]

China Inf

Liang, D.et al.Relative difficulty distillation for semantic segmentation.Sci. China Inf. Sci.67, 192105, DOI: 10.1007/ s11432-023-4012-0 (2024)

2024
[22]

Dong, W., Du, B. & Xu, Y . Shape-intensity knowledge distillation for robust medical image segmentation.Front. Comput. Sci.19, 199705, DOI: 10.1007/s11704-024-4066-4 (2025)

work page doi:10.1007/s11704-024-4066-4 2025
[23]

Hu, M.et al.Knowledge distillation from multi-modal to mono-modal segmentation networks.Int. Conf. on Med. Image Comput. Comput. Interv. (MICCAI)553–562, DOI: 10.1007/978-3-030-59728-3_54 (2020). 25.Ma, J.et al.Segment anything in medical images.Nat. Commun.15, 654 (2024)

work page doi:10.1007/978-3-030-59728-3_54 2020
[24]

& Efros, A

Zhu, J.-Y ., Park, T., Isola, P. & Efros, A. A. Unpaired image-to-image translation using cycle-consistent adversarial networks. InProceedings of the IEEE International Conference on Computer Vision (ICCV), 2223–2232 (2017). 13/14

2017
[25]

InProceedings of the IEEE international conference on computer vision, 2794–2802 (2017)

Mao, X.et al.Least squares generative adversarial networks. InProceedings of the IEEE international conference on computer vision, 2794–2802 (2017)

2017
[26]

& Kweon, I

Woo, S., Park, J., Lee, J.-Y . & Kweon, I. S. Cbam: Convolutional block attention module. InProceedings of the European conference on computer vision (ECCV), 3–19 (2018)

2018
[27]

& Adam, H

Chen, L.-C., Zhu, Y ., Papandreou, G., Schroff, F. & Adam, H. Encoder-decoder with atrous separable convolution for semantic image segmentation. InProceedings of the European conference on computer vision (ECCV), 801–818 (2018)

2018
[28]

Chen, J.et al.Transunet: Transformers make strong encoders for medical image segmentation.arXiv preprint arXiv:2102.04306(2021)

work page internal anchor Pith review arXiv 2021
[29]

InProceedings of the IEEE/CVF winter conference on applications of computer vision, 574–584 (2022)

Hatamizadeh, A.et al.Unetr: Transformers for 3d medical image segmentation. InProceedings of the IEEE/CVF winter conference on applications of computer vision, 574–584 (2022)

2022
[30]

F., Kohl, S

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.Nat. methods18, 203–211 (2021). Author contributions J.P. developed the algorithm, conducted the experiments, and wrote the initial manuscript. J.C. revised the manuscript. J.P.Y . proposed the ...

2021