PanoDiff-SR: Synthesizing Dental Panoramic Radiographs using Diffusion and Super-resolution

arXiv:2507.09227 · eess.IV, cs.AI, cs.CV, cs.LG · Submitted 2025-07-12 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "PanoDiff-SR: Synthesizing Dental Panoramic Radiographs using Diffusion and Super-resolution".

Jane: The gist: This study proposes PanoDiff, a combination of diffusion-based generation and Super-Resolution,

Tom: First, who's behind it and why it matters.

Title and authors: Jane: So, what does the actual summary say about how PanoDiff works? Can you break down that technical description for us?

Tom: It says the core idea is to sample from a manifold representation using a trained sampler G theta. This sampler takes noise, which is normally sampled from N(zero I), and transforms it into high-quality images that match the target distribution p(x) > <ref:2507.09227#pg3>.

Lu: They use a U-Net architecture with hidden layers of dimensions sixty-four one hundred twenty-eight two hundred fifty-six five hundred twelve. The self-attention mechanism in the U-Net is specifically useful here because it lets the model capture long-range dependencies across the entire panoramic radiograph.

Jane: Long-range dependencies are important for anatomy. Does that mean it’s better at keeping structures in place during generation?

Tom: Exactly. The self-attention allows it to ensure that anatomical structures remain spatially coherent during generation >

Meng: That sounds like a lot of computation just to keep things organized across the whole image. What about the noise scheduling? How do they make sure the noise process is smooth and stable?

Lalam: They use a cosine-based beta schedule for adding noise, which ensures smoother transitions between noise levels, leading to more stable training and higher quality outputs compared to a linear schedule >

The paper's summary: Jane: Beyond just describing how it works, what are the specific improvements they claim they have over existing methods? What makes PanoDiff-SR better than just using a standard generator followed by a basic upscaler?

Tom: They propose several things. First, the diffusion model itself uses that cosine-based beta schedule we just talked about, and during inference, they use a DDIM scheduler instead of the original T = one thousand steps <ref:2507.09227#pg1>.

Lu: That DDIM scheduler reduces the number of required denoising steps from T = one thousand down to T = two hundred fifty which is a significant reduction in computational time for generating an image > <ref:2507.09227#pg1>

Meng: Two thousand five hundred instead of one thousand? That’s a lot less compute time for inference, which would help if you were deploying this on a system with limited resources. But what about the super-resolution part?

Jane: The super-resolution module uses a Transformer-based architecture with hybrid attention mechanisms, specifically window-based multi-head self-attention and overlapping cross-attention >

Tom: That design allows it to capture both local and global dependencies in the image, which is key. It lets the model focus on relevant regions of the image to reconstruct fine anatomical details >

Lu: Plus, they use residual hybrid attention groups, or RHAGs, which enables the model to progressively refine features across multiple layers for better quality >

Meng: And they train it using a hybrid loss function defined as Ltotal = lambda1 L1 + lambda2 Lpercep + lambda3 LGAN. So it’s not just one loss trying to do everything >

The paper's improvements: Jane: So, to wrap this up, what is the overall meaning of PanoDiff-SR for the field right now? What’s the big picture here?

Tom: It means they have a robust way to generate synthetic dental panoramic radiographs that look very realistic. They're using diffusion and super-resolution together to create high-quality outputs.

Lu: The implication is that we can create large, high-fidelity synthetic datasets for training AI models in dentistry without relying solely on limited real data >

Meng: For deployment, the trade off seems to be complexity. You have a complex diffusion model and a transformer SR module, but if the output quality is high enough for your specific use case, that complexity might be worth it >

Lalam: They also did this evaluation by having clinical experts look at a mixture of one hundred synthetic and one hundred real PRs <ref:2507.09227#pg1>. The average accuracy in distinguishing real from synthetic was sixty-eight point five percent, with fifty percent corresponding to random guessing >

Jane: That human study result is important because it shows that the images are quite good at fooling people, which speaks to the quality of the generation itself.

Tom: It's solid performance when you look at the technical metrics too. They got a Fréchet inception distance score of forty point six nine between seven thousand two hundred forty-three real and synthetic images in HR >

Lu: And they also showed high scores for LRPD and HRPD FID scores, which confirms that the synthetic data is statistically similar to the real data in terms of distribution >

Meng: So, the system is quite good at matching the distribution, but we still have to think about how to make it work reliably when you're trying to deploy it in a clinical setting.

Jane: Exactly. We’ve got a very high-quality synthetic image generation method for dental X-rays here with PanoDiff-SR >

Conclusion: Tom: So we’re wrapping up on "PanoDiff-SR: Synthesizing Dental Panoramic Radiographs using Diffusion and Super-resolution". Basically, they showed how you can use diffusion and super-resolution together to make really realistic fake dental X-rays.

Jane: Right, so it’s taking a small image of size two hundred fifty-six by one hundred twenty-eight and turning it into a high-quality picture of size one thousand twenty-four by five hundred twelve using this two-step process. It sounds like they're solving the problem of getting realistic synthetic data for training AI in dentistry.

Lu: Exactly. The core idea is using diffusion to sample from a distribution that matches real dental X-rays, and then the super-resolution module uses a hybrid attention transformer to reconstruct those fine anatomical details properly.

Meng: I still have some questions about the deployment side though. Generating that one thousand twenty-four by five hundred twelve image from a seed sounds computationally heavy for running this in production environments.

Lalam: It's designed to generate images that look realistic enough to be useful for training, and they’re using a U-Net architecture enhanced with self-attention mechanisms and residual blocks in the diffusion model part of PanoDiff.

Jane: I see. So what about the training setup? How do they make sure the generation process is stable when it's running?

Tom: They use a cosine-based beta schedule for adding noise, which ensures smoother transitions between noise levels, leading to more stable training and higher quality outputs compared to a linear schedule.

Lu: And during inference, they use a DDIM scheduler instead of the original one thousand steps. That reduces the number of required denoising steps from one thousand down to two hundred fifty which is a significant reduction in computational time for generating an image.

Jane: The super-resolution module uses a Transformer-based architecture with hybrid attention mechanisms, specifically window-based multi-head self-attention and overlapping cross-attention. That design allows it to capture both local and global dependencies in the image.

Tom: That lets the model focus on relevant regions of the image to reconstruct fine anatomical details, which is key for X-rays. Plus, they use residual hybrid attention groups, or RHAGs, which enable the model to progressively refine features across multiple layers for better quality.

Lu: They also train it using a hybrid loss function defined as Ltotal equals lambda1 L1 plus lambda2 Lpercep plus lambda3 LGAN. So it’s not just one loss trying to do everything; they're balancing accuracy, perceptual realism, and adversarial training.

Lalam: They also did this evaluation by having clinical experts look at a mixture of one hundred synthetic and one hundred real PRs. The average accuracy in distinguishing real from synthetic was sixty-eight point five percent, with fifty percent corresponding to random guessing.

Jane: That human study result is important because it shows that the images are quite good at fooling people, which speaks to the quality of the generation itself.

Lu: And they also showed high scores for LRPD and HRPD FID scores, which confirms that the synthetic data is statistically similar to the real data in terms of distribution.

Meng: So the system is quite good at matching the distribution, but we still have to think about how to make it work reliably when you're trying to deploy this in a clinical setting.

Jane: Exactly. We’ve got a very high-quality synthetic image generation method for dental X-rays here with PanoDiff-SR. It’s a big step toward making AI training data more accessible, but what happens next?

Department of Dentistry and Oral Health, Aarhus University · Aarhus Institute of Advanced Studies, Aarhus University · Department of Mathematics, Aarhus University · Faculty of Information Technology and Communication Sciences, Tampere University

eess.IV, cs.AI, cs.CV, cs.LG

Submitted: 2025-07-12

Updated: 2026-10-15

Comments: 28 pages, 19 figures, 11 tables. Project page: https://s4nyam.github.io/panodiff/ Code: https://github.com/s4nyam/PanoDiff

Code: https://github.com/s4nyam/PanoDiff

Project page: https://s4nyam.github.io/panodiff

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 90/100

The gist: The gist: This study proposes PanoDiff, a combination of diffusion-based generation and Super-Resolution, to synthesize high-quality dental panoramic radiographs (PRs) by generating low-resolution

Key concepts

PanoDiff Architecture: This is the overall system that works in two main steps: first, a diffusion model creates low-resolution (256x128) image seeds by gradually adding noise. Second, a Super-Resolution module uses a transformer to upscale these seeds into high-resolution (1024x512) images. This combination ensures both realistic generation and fine detail reconstruction.
Diffusion Model: A diffusion model generates images by learning to reverse a process of adding Gaussian noise. It starts with pure noise and iteratively denoises it over many steps, guided by a learned prediction network. In PanoDiff, this process is used to create the initial low-resolution image structure before the Super-Resolution step refines the details.
Hybrid Attention Transformer (HAT): This is the core of the Super-Resolution module. It uses a transformer architecture with hybrid attention mechanisms—specifically window-based multi-head self-attention and overlapping cross-attention. This allows the model to simultaneously look at local image features and global context, which is crucial for accurately reconstructing fine anatomical details in dental X-rays.

Terminology

Accuracy — 1× in this paper · explained in 1 episode
Transformer — 1× in this paper · explained in 2 episodes

Summary

The gist: This study proposes PanoDiff, a combination of diffusion-based generation and Super-Resolution, to synthesize high-quality dental panoramic radiographs (PRs) by generating low-resolution seeds and then upscaling them using a transformer model.

PanoDiff Architecture

PanoDiff is designed as a diffusion model that generates low-resolution PRs of size 256 × 128, which are subsequently processed by an SR module to yield high-resolution (HR) PRs of size 1024 × 512 The top-level principle of PanoDiff, illustrated in Figure 3, involves generating synthetic images by sampling from a manifold representation using a trained sampler Gθ that transforms noise z ∼ N(0, I) into high-quality images resembling the target distribution p(x) The model employs a U-Net architecture enhanced with self-attention mechanisms and residual blocks The forward process gradually adds Gaussian noise to the original image x0 over T time steps, producing a noisy image xt as described in Equation 1 The reverse process, guided by a learned noise prediction network ϵθ(xt, t), iteratively denoises xt to recover x0

Diffusion Model Details

The diffusion model in PanoDiff utilizes a U-Net architecture consisting of hidden layers with dimensions [64, 128, 256, 512], which progressively downsample and upsample the input image while maintaining feature richness at each resolution level The self-attention mechanism integrated into the U-Net is particularly useful in PanoDiff, as it allows the model to capture long-range dependencies across the PR, ensuring that anatomical structures remain spatially coherent during generation PanoDiff uses a cosine-based beta schedule for noise addition, defined in Equation 2 This schedule ensures smoother transitions between noise levels, leading to more stable training and higher-quality outputs compared to the linear schedule used originally in [23] Furthermore, the model employs a DDIM scheduler during inference, which reduces the number of required denoising steps from T = 1000 to T = 250 Another novel aspect of PanoDiff is the use of an exponential moving average (EMA) with a cosine-based update schedule for model weights

Super-Resolution Module

The super-resolution (SR) module in our framework is designed to enhance low-resolution PRs to high-resolution images of size 1024 × 512 Unlike traditional SR methods that rely solely on convolutional neural networks (CNNs), our approach leverages a Transformer-based architecture with hybrid attention mechanisms, built on a Hybrid Attention Transformer (HAT) [2] The key innovation of HAT lies in the integration of window-based multi-head selfattention and overlapping cross-attention, which capture both local and global dependencies in the image This allows the model to focus on relevant regions of the image, improving the reconstruction of fine anatomical details Additionally, the use of residual hybrid attention groups (RHAG) enables the model to progressively refine features across multiple layers, enhancing the overall quality of the upscaled images To simulate real-world degradations, we employ a novel two-stage degradation process to create HR-LR image pairs

Training and Evaluation

PanoDiff was trained for 110 epochs using 7243 training images of PRs (H = 512, W = 1024) For SR, while we propose new methods and losses, we used the rest of the implementation from HAT, thus using the pretrained weights for the generator released for Real HAT GAN SR×4 in [2, 3] and fine-training it with our dataset’s image pairs The training process involves two degradation stages with various augmentations, including resizing, noise addition, and blurring The total loss for the generator is defined in Equation 8 This hybrid approach, combining transformer-based attention mechanisms with GAN training and multi-loss optimization, has been shown to significantly outperform traditional SR methods [3]

Experimental Results

Experimental results demonstrate a Fréchet inception distance score of 40.69 between 7243 real and synthetic images (in HR) Inception scores were 2.55, 2.30, 2.90 and 2.98 for real HR, synthetic HR, real LR and synthetic LR images, respectively Among a diverse group of six clinical experts, all evaluating a mixture of 100 synthetic and 100 real PRs in a time-limited observation, the average accuracy in distinguishing real from synthetic images was 68.5% (with 50% corresponding to random guessing) The inter-mixing of the data points (clusters) in the t-SNE plots from Figure 6, the low HRGT vs HRPD FID score of 40.69 in Table 3, high IS for LRPD and HRPD in Table 4 as well as the results from The clinical expert evaluation from Table 5 show the performance of our proposed method

Conclusion

In this study, we optimized diffusion methods (PanoDiff) and SR methods to synthesize high-quality realistic PRs Furthermore, apart from more technical evaluation methods like t-SNE, FID, and IS, we also performed a human observer study involving dentists to distinguish real from synthetic PRs Future work can focus on conditionally synthesizing lesions or other pathological conditions within the PRs (e.g. Table 5) The clinical expert evaluation from Table 5 show the performance of our proposed method The inter-mixing of the data points (clusters) in the t-SNE plots from Figure 6, the low HRGT vs HRPD FID score of 40.69 in Table 3, high IS for LRPD and HRPD in Table 4 as well as the results from The clinical expert evaluation from Table 5 show the performance of our proposed method The inter-mixing of the data points (clusters) in the t-SNE plots from Figure 6, the low HRGT vs HRPD FID score of 40.69 in Table 3, high IS for LRPD and HRPD in Table 4 as well as the results from The clinical expert evaluation from Table 5 show the performance of our proposed method The inter-mixing of the data points (clusters) in the t-SNE plots from Figure 6, the low HRGT vs HRPD FID score of 40.69 in Table 3, high IS for LRPD and HRPD in Table 4 as well as the results from The clinical expert evaluation from Table 5 show the performance of our proposed method The inter-mixing of the data points (clusters) in the t-SNE plots from Figure 6, the low HRGT vs HRPD FID score of 40.69 in Table 3, high IS for LRPD and HRPD in Table 4 as well as the results from The clinical expert evaluation from Table 5 show the performance of our proposed method The inter-mixing of the data points (clusters) in the t-SNE plots from Figure 6, the low HRGT vs HRPD FID score of 40.69 in Table 3, high IS for LRPD and HRPD in Table 4 as well as the results from The clinical expert evaluation from Table 5 show the performance of our proposed method The inter-mixing of the data points (clusters) in the t-SNE plots from Figure 6, the low HRGT vs HRPD FID score of 40.69 in Table 3, high IS for LRPD and HRPD in Table 4 as well as the results from The clinical expert evaluation from Table 5 show the performance of our proposed method The inter-mixing of the data points (clusters) in the t-SNE plots from Figure 6, the low HRGT vs HRPD FID score of 40.69 in Table 3, high IS for LRPD and HRPD in Table 4 as well as the results from The clinical expert evaluation from Table 5 show the performance of our proposed method The inter-mixing of the data points (clusters) in the t-SNE plots from Figure 6, the low HRGT vs HRPD FID score of 40.

Improvements for AI systems

The PanoDiff diffusion model can generate low-resolution PR seeds (256 × 128) which are then processed by a state-of-the-art transformer to yield high-resolution (HR) PRs of size 1024 × 512, thus providing a pathway for generating high-quality synthetic dental panoramic radiographs.
The SR module uses a transformer-based SR network that captures local and global dependencies to reconstruct sharper edges and finer textures, improving the depiction of anatomical details compared to conventional CNN-based methods.
The system can be trained using a hybrid loss function, defined as Ltotal = λ1L1 + λ2Lpercep + λ3LGAN, which combines pixel-wise accuracy (pixel-wise loss), visual realism (perceptual loss), and indistinguishability from real images (adversarial loss).
The SR module employs a novel two-stage degradation process to create HR-LR image pairs involving Poisson noise to the ground-truth images, followed by JPEG compression, which allows the SR model to handle various image quality manifestations.
The PanoDiff model incorporates advanced training techniques such as a cosine beta schedule for noise addition and a DDIM scheduler during inference, which significantly reduce the number of sampling steps from T = 1000 to T = 250 while maintaining high image quality.
The system can be evaluated by conducting human observer studies where dentists distinguish between real and synthetic PRs, with results showing that EPs (average AUC = 0.87) performing better than ECs (average AUC = 0.79) at distinguishing real from fake PRs.
A custom ViT classifier trained on HR GT and HR PD images can be used to visualize attention maps, which may reveal that for high-realism synthetic images, the ViT’s AMs began to resemble those seen in real images, serving as an implicit evaluation of the diffusion model’s generative quality.