PanoDiff-SR: Synthesizing Dental Panoramic Radiographs using Diffusion and Super-resolution
Authors: Sanyam Jain, Bruna Neves de Freitas, Andreas Basse-OConnor, Alexandros Iosifidis, Ruben Pauwels
Organizations: Department of Dentistry and Oral Health, Aarhus University, 8000 Aarhus, Denmark · Aarhus Institute of Advanced Studies, Aarhus University, 8000 Aarhus, Denmark · Department of Mathematics, Aarhus University, 8000 Aarhus, Denmark · Faculty of Information Technology and Communication Sciences, Tampere University, Tampere, Finland
Interest in generating realistic synthetic medical images has grown in recent years. Such datasets can mitigate the scarcity of public data for artificial intelligence research and support education. Work on synthesizing dental panoramic radiographs (PRs) has so far relied almost exclusively on generative adversarial networks (GANs), and three limitations recur: the synthesized field of view is often restricted to a cropped dentoalveolar region rather than a complete PR; observer studies of GAN-generated PRs report that clinicians still identify synthetic images with accuracies of roughly 78-82%, implying persistent visible giveaways; and adversarial training remains sensitive to instability and mode collapse. Generating a full-size PR directly at diagnostic resolution with a diffusion model is an alternative, but sampling at that resolution is costly. We therefore present PanoDiff-SR, an end-to-end two-stage pipeline developed and evaluated specifically for panoramic radiography: a denoising diffusion implicit model (PanoDiff) generates a low-resolution (LR) seed PR (256×128), which a fine-tuned hybrid attention transformer (HAT) super-resolution (SR) stage upscales to a high-resolution (HR) PR of 1024×512. Decoupling the expensive generative step from the resolution step yields complete, full-arch PRs at diagnostic resolution on a single 47 GB GPU. The Fréchet inception distance between 7243 real and synthetic HR images is 40.5. Inception scores were 2.68, 2.38, 2.90 and 2.98 for real HR, synthetic HR, real LR and synthetic LR images, respectively. Six clinical experts, each evaluating 100 synthetic and 100 real PRs under a time limit, distinguished real from synthetic images with an average accuracy of 68.5% (50% corresponds to random guessing), substantially closer to chance than reported for GAN-based studies. Project page: https://s4nyam.github.io/panodiff/
Figures & tables
Fig. 1: Selected synthetic examples from our work that fooled most dentists.
Abbr.
Images
Country
Resolution
ADLD [ 36 ]
500*
China
∼ 2940 × 1435 or ∼ 987 × 478
DENTEX [ 37 ]
3603
Switzerland
∼ 2950 × 1316 or ∼ 1976 × 976
TSXK [ 38 ]
598 ( ×2 )
DR Congo
2041 × 1024
TUFTS [ 39 ]
1000
USA
1615 × 840
USPFORP [ 40 ]
945
Brazil
2903 × 1536
TABLE I: Overview of dental radiography datasets used in our study. Dataset abbreviations with format, year, and availability: ADLD (png, 2024, Kaggle)*, DENTEX (png, 2023, Zenodo), TSXK (png, 2023, Kaggle), TUFTS (jpg, 2022, On Request), USPFORP (jpg, 2024, On Request). * indicates that the dataset was recently updated with 1500 more images, but we accessed it when it had 500 images. The ’Images’ column denotes the radiograph files used from each release. ∼ indicates varying sizes in the dataset within the given resolution range. Abbreviations: ADLD – A dual-labeled dataset, DENTEX – Dental Enumeration and Diagnosis on Panoramic X-rays, TSXK – Teeth Segmentation on dental X-ray images, TUFTS – Tufts Dental Database, USPFORP – São Paulo dataset.
Fig. 2: T-distributed stochastic neighbor embedding (t-SNE) plot of 500 random images picked from each source dataset.
Fig. 3: General principle of synthetic image generation through manifold representation. Consider a dataset of images {xk}k=1n , where xk∼p(x) . These images serve as samples from the target distribution p(x) . A best sampler Gθ is one such that x^=Gθ(z) , where z∼N(0,I) , to produce high-quality samples resembling the true data distribution p .
Fig. 4: Forward phase of PanoDiff. Noise is added to the input image xt=0 over T=1000 time steps following the cosine β -schedule, until the image is pure noise at t=1000 . The plot at the lower right tracks the pixel statistics along the trajectory.
Fig. 5: Reverse phase of PanoDiff. Top: the U-Net (flame: trainable) is trained on (x0,xt=k) with an L1 loss to predict the noise contained in xt=k . Bottom, for comparison: step-by-step DDPM denoising with the frozen network (snowflake), for which a single-step estimate is not adequate at high noise levels.
Fig. 6: Image generation in PanoDiff. Starting from noise, the frozen U-Net predicts and removes noise, noise is added back, and the loop repeats for the number of inference time steps. Sixteen generated PRs are shown at the lower right.
PanoDiff
SR (HAT, 4× )
(LR synthesis)
(LR → HR)
Trainable parameters
34.0 M
20.8 M
Training input
256×128
256×256 crops
Batch size
4
4
Training schedule
110 epochs
400,000 iter.
(199,210 steps)
( ≈ 221 epochs)
TABLE II: Computational cost of the two stages of PanoDiff-SR. The training figures refer to a single NVIDIA RTX 6000 Ada Generation GPU (47 GB); the inference times were measured for both stages under identical conditions on the GPU used in Section IV-E2 . LR: 256×128 ; HR: 1024×512 . The SR stage is fine-tuned on random 256×256 crops of the real HR radiographs, the standard practice for HAT; since it accepts any image size, it is applied to the whole 256×128 output of PanoDiff at inference, without padding or rescaling. Each schedule is given in the unit of its training code, with the equivalent in the other unit in brackets.
Fig. 7: Comparison of generated PRs across epochs. Each column represents a different epoch from left to right, showing the images generated using same unique seed per row.
Fig. 8: Visualization of t-SNE embeddings in 2D for low-resolution (LR, left) and high-resolution (HR, right) data for ground truth (GT, i.e. real data) and PanoDiff (PD, i.e. synthetic data). Euclidean distances between the centroids of these clusters are - LR: 81.41, HR: 34.73. To keep the plot readable with 2×7243 points per panel, a random subsample of 1200 points per class is drawn as a scatter, while the contours give the kernel density of the complete set at 25%, 50% and 75% of its peak; crosses mark the centroids of the complete sets and the dashed line connects them.
Comparison Type
Combination
FID
Self-comparisons
HRGT-HRGT
5.7
LRGT-LRGT
6.2
HRPD-HRPD
5.0
LRPD-LRPD
5.9
Real vs. Synthetic
LRGT-LRPD
55.0
HRGT-HRPD
40.5
TABLE III: Comparison of real vs. synthetic images using FID. Lower scores indicate greater similarity.
Source
IS Mean
IS Std
Source
IS Mean
IS Std
HRGT
2.684
0.064
ADLD
2.938
0.037
HRPD
2.381
0.031
DENTEX
2.905
0.051
LRGT
2.898
0.042
TSXK
2.644
0.053
LRPD
2.983
0.066
TUFTS
2.627
0.047
G
1.050
0.001
USPFORP
2.882
0.034
TABLE IV: IS, with higher values indicating more diversity and image quality. The left side of the table shows IS for {LR, HR} + {GT, PD}, the right side shows IS for the datasets of real images used in this study. G corresponds to pure Gaussian noise images.
100 synthetic PRs
100 real PRs
Metrics
TLD
DR
PR
U
PF
DF
DR
PR
U
PF
DF
P
R
A
EC1
3
7
16
34
40
13
20
32
25
10
0.60
0.75
0.63
EC2
4
22
3
25
46
26
46
3
19
6
0.68
0.72
0.69
EC3
0
4
11
45
40
0
38
33
28
1
0.63
0.80
0.66
EP1
0
6
7
83
4
1
54
33
12
0
0.65
0.71
0.66
EP2
2
12
0
33
53
40
44
0
16
0
0.78
0.81
0.79
TABLE V: Observer responses and derived performance metrics for the six dentists. The two left-hand blocks give the raw number of responses in each category of the five-level scale, separately for the 100 synthetic and the 100 real PRs; every row of each block sums to 100 and no weighting of any kind has been applied. DR: definitely real; PR: probably real; U: unsure; PF: probably fake; DF: definitely fake. The right-hand block gives precision, recall and accuracy computed from the certainty-weighted counts as described in the text.
Fig. 9: Pie charts for each observer showing distribution of correct and incorrect decisions. ‘Fully’ and ‘partially’ refers to the level of certainty indicated by the observer for a given answer, as described in the text.
Fig. 10: Left: ROC plot for individual ECs and their average. Right: ROC plot for individual EPs and their average. The value in brackets after each label is the AUC.
Fig. 11: Left: precision–recall curve for individual ECs and their average. Right: precision–recall curve for individual EPs and their average. The value in brackets after each label is the average precision (AP).
Fig. 12: Examples of PRs from expert evaluation. Within each category, the first row corresponds to fully certain (FC) and the second row to partially certain (PC) responses. Examples were selected for each category based on the majority of the observers’ assessments.
EC1
EC2
EC3
EP1
EP2
EP3
EC1
1
0.42
0.50
0.44
0.47
0.53
EC2
0.42
1
0.56
0.45
0.49
0.49
EC3
0.50
0.56
1
0.53
0.50
0.62
EP1
0.44
0.45
0.53
1
0.57
0.47
EP2
0.47
0.49
0.50
0.57
1
0.50
EP3
0.53
0.49
0.62
0.47
0.50
1
TABLE VI: Weighted Cohen’s Kappa values between observers.
Fig. 13: Attention maps generated using a trained ViT for two classes - top three rows for real images and bottom three rows for synthetic images. The confidence value ‘C’ represents the ViT classifier’s output and ranges from 0 (for synthetic) to 1 (for real). Prediction values in red correspond to incorrect classification.
Fig. 14: Attention statistics. Top: a held-out radiograph (1), the corresponding attention-rollout map from the real-versus-synthetic ViT (2), and the fixed anatomical templates (3), comprising the TUFTS teeth masks (dashed) and maxillomandibular masks (solid), averaged and thresholded at 0.5 . For every image, the share of attention inside the teeth template, inside the jaw template, and the entropy of the map were calculated. Bottom (4): distributions of those numbers over the 1449 real and 1449 synthetic held-out radiographs. The entropy separates the two classes, whereas the anatomical shares overlap considerably.
Fig. 15: MedSAM with the dentition-box prompt on real radiographs, three from each labelled device, and on PanoDiff-SR radiographs. The yellow box is the prompt, drawn around the U-Net tooth region in the same way on both sets; the cyan contour is MedSAM’s mask; the badge is its predicted mask quality.
Fig. 16: t-SNE of every generative arm in the Inception feature space used for all FID and IS values, on 1500 images per arm; each panel is embedded separately (PCA to 50 dimensions, then t-SNE with perplexity 30). Numbers in parentheses are the source indices of Table VIII . Crosses mark cluster centroids and lines enclose half of each arm’s peak density. Distances between clusters in a t-SNE map are not quantitatively meaningful, but measured in the original feature space, the distance from each full-resolution generator in panel (c) to the real images orders the arms exactly as their FIDs in Table X do.
Index
Source
IS Mean
IS Std
1
LRGT
2.898
0.042
2
LRPD
2.983
0.066
3
LRGAN
2.253
0.034
4
HRPD
2.381
0.031
5
LRGAN+HATSR
2.047
0.020
6
LRPD+SwinIR
2.887
0.073
TABLE VIII: IS for different sources, each of 7243 images; the index is used by Table IX and Figures 16 and 17 . HAT-SR and SwinIR were both fine-tuned on the 7243 real PRs. LR: low-resolution, HR: high resolution. GT: ground truth (i.e., real images), PD: PanoDiff diffusion, GAN: generative adversarial network.
Fig. 17: Precision (fidelity) and recall (coverage) for (a) the two generators at low resolution and (b) the four combinations at full resolution. Every image has a neighbourhood in the Inception feature space: the region within the distance to its third-nearest neighbour in its own set. Top row: for each generated image, its smallest distance to a real image, measured in units of that real image’s neighbourhood radius; images left of the dashed line lie inside the neighbourhood of at least one real image, and their share, printed in each panel in the colour of its curve, is the precision. Bottom row: the same for each real image with respect to the generated images; the share inside is the recall. Curves represent kernel density estimates over all 7243 images of each set. Numbers in parentheses are the source indices of Table VIII .
Fig. 18: Low-resolution synthesis and super-resolution for three samples of each generator (columns). Each column traces one seed from the 256×128 sample, upsampled bicubically for display, through HAT-SR and SwinIR, for PanoDiff (top three rows) and FastGAN (bottom three rows). The inset in every panel magnifies the region outlined in yellow.
#
Model
FID ↓
IS ↑
Prec. (%) ↑
Rec. (%) ↑
ViT det. (%) ↓
Cost (GPU-h)
1
PanoDiff-SR (ours)
40.50
2.38
56.2
28.8
97.6
82 †
2
FastGAN + HAT-SR [ 55 , 20 ]
102.27
2.05
6.9
0.8
99.9
64 †
3
StyleGAN2-ADA [ 63 ]
23.45
2.58
61.1
21.9
89.6
441
4
StyleGAN3 [ 29 ]
19.38
2.66
67.3
25.5
77.9
572
5
FastGAN-HR [ 55 ]
215.89
2.30
0.0
0.0
100.0
11
6
ADM [ 16 ]
65.03
2.71
28.2
16.6
98.4
122
TABLE X: Direct high-resolution synthesis compared with the two-stage pipeline. Rows 1 and 2 generate at 256×128 and are upscaled to 1024×512 by HAT-SR; rows 3–8 generate 1024×512 directly and were trained from randomly initialised weights on the same 7243 real PRs with comparable iterations and training budget. FID and IS are computed against the same 7243 real HR PRs. “ViT det.” is the held-out accuracy of the real-versus-synthetic classifier of Section IV-C retrained per arm, so lower means harder to tell from a real radiograph. Precision and recall are computed as in Table IX and Figure 17 ( k=3 ). Training cost is in GPU-hours, all on the same type of GPU. † Measured time per training step of each stage multiplied by its number of training steps; the SR stage accounts for 61.5 of these hours in both rows.
Fig. 19: Selected examples of the image sets of Table XI : real radiographs, the two-stage pipeline (PanoDiff-SR) and the direct high-resolution models of Table X .
Bilateral
Tooth-row
Occlusal
Outside
Images
symmetry r
periodicity
separation
real range
Real
0.477±0.199
0.133±0.116
0.503±0.283
14.4%
PanoDiff-SR
0.276±0.172∗∗
0.169±0.124∗∗
0.571±0.259∗∗
20.1%
StyleGAN2-ADA
0.562±0.188∗∗
0.113±0.108∗∗
0.498±0.272
16.2%
StyleGAN3
0.491±0.195∗
0.111±0.113∗∗
0.500±0.278
17.0%
ADM
0.178±0.145∗∗
0.161±0.110∗∗
0.475±0.275∗
27.1%
TABLE XI: Anatomical plausibility measures, on 1500 images per model. The three measures are given as mean ± standard deviation; ∗ and ∗∗ mark a difference from the real images at p<0.05 and p<10−4 (Mann–Whitney U ). The last column is the fraction of images falling outside the interval spanned by the central 95% of real radiographs on any of the three measures, so the first row shows what a generator that matches the real distribution would score.
We introduce the first generative foundation model for chest radiograph synthesis trained from scratch at the billion-parameter scale. Existing radiographic AI models often suffer from poor generalisation across patient subpopulations, institutions, and acquisition settings, resulting in limited real-world clinical utility. Controlled, high-fidelity synthesis of chest radiographs is a promising path toward diversifying clinical datasets and evaluating the robustness of diagnostic models. Therefore, we present the largest specialist generative foundation model for chest radiographs to date, with over 1.3B parameters, trained for 1.6T tokens on a curated, heterogeneous dataset comprising 1.2M radiographs and clinical expert-guided metadata. Our model supports controllable radiograph generation and editing across multiple demographic subgroups, acquisition views, and a dozen pathologies. Moreover, we significantly advance the state of the art in radiograph synthesis fidelity, producing images that are indistinguishable from real radiographs to clinical experts.
Fabio De Sousa Ribeiro, Emma A. M. Stanley, Charles Jones +7
Imperial College London · Causality in Healthcare AI Hub · Cleveland Clinic London +4
Panoramic radiograph (PR) is fundamentally used in routine dental care, but it inherently provides only a two-dimensional (2D) projection of complex three-dimensional (3D) craniofacial anatomy. Most existing learning-based methods attempt to computationally recover this 3D information by directly regressing native cone-beam computed tomography (CBCT) volumes from PR. However, this direct mapping requires the model to simultaneously learn common anatomical structures and patient-specific morphological variations. This entangled formulation makes the ill-posed 2D-to-3D inverse problem highly ambiguous, often producing over-smoothed reconstructions with blurred anatomical boundaries. To address this, we propose HyDAR-Pano3D, a two-stage framework that reformulates PR-to-CBCT reconstruction as a disentangled anatomical recovery problem. In Stage 1, a dual-encoder network integrates radiographic features with SAM-derived semantic priors to reconstruct an arch-normalized canonical volume. In Stage 2, an Anatomical Restoration Network predicts a prior-constrained structured deformation field to map this canonical volume back to the native space, restoring individual morphological variations. Experiments on three large-scale datasets show that HyDAR-Pano3D significantly outperforms baseline methods (p<0.05), achieving a 25.76 dB PSNR, 85.70% SSIM, and an 83.83% overall anatomical Dice score. The synthesized volumes successfully support downstream segmentation of whole teeth (82.4% Dice) and the inferior alveolar canal (72.2% Dice), demonstrating that our disentangled approach preserves clinically relevant structures to enable robust anatomy-aware assessment when CBCT data is unavailable.
Yaoyao Yue, Jérôme Schmid, Xiaoshuang Li +2
School of Computer Science, the University of Sydney, Australia · Geneva School of Health Sciences, HES-SO University of Applied Sciences and Arts Western Switzerland, Geneva, Switzerland · Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai 200240, China +1
Panoramic dental radiographs require fine-grained spatial reasoning, bilateral symmetry understanding, and multi-step diagnostic verification, yet existing vision-language models operate under a static single-pass paradigm that limits their clinical reliability. In this paper, we introduce OralGPT-Plus, an agentic vision-language model designed to perform iterative and symmetry-aware diagnostic reasoning for panoramic dental radiograph analysis. To support this paradigm, we construct DentalProbe, a five-thousand-image dataset with expert-curated diagnostic trajectories that provide structured supervision for localized inspection and contralateral comparison. We further develop a Reinspection-driven reinforcement learning framework that encourages clinically meaningful re-examination and stabilizes long-horizon reasoning with rubric-based reward and conditioned diagnostic-driven reward. In parallel, we present MMOral-X, the first benchmark for holistic panoramic diagnosis, containing 300 open-ended questions and region-level annotations across multiple difficulty levels. OralGPT-Plus demonstrates consistent and reliable improvements over strong baselines on MMOral-X and established panoramic benchmarks, indicating the effectiveness of interactive and symmetry-informed reasoning. Our work highlights the value of agentic modeling for dental imaging and provides a foundation for future research in clinically aligned panoramic radiograph analysis.
Yuxuan Fan, Jing Hao, Hong Chen +5
The Hong Kong University of Science and Technology (GZ) · School of Computer Science, Peking University · Faculty of Dentistry, The University of Hong Kong +3