Background: Cone-beam computed tomography (CBCT) is a crucial tool for visualizing alveolar bone defects, such
as fenestrations and dehiscences, in periodontal imaging. However, there aren't many publicly available CBCT
datasets because of privacy concerns, cost, and radiation exposure. This makes it challenging to develop robust AIdriven diagnostic models. Generative models like GANs don't work as well with small amounts of data, which is why we need better, more stable, and data-efficient options. This study proposes a comprehensive in silico framework that utilizes a hybrid 3D diffusion-autoencoder model to generate synthetic CBCT volumes of alveolar bone defects,
eliminating the need for real training data.
Methods: We used a two-stage generative model. A 3D convolutional autoencoder compressed 64³ voxel patches into
a 256-dimensional latent space. Then, we trained a denoising diffusion probabilistic model (DDPM) on latent vectors
with added noise to produce realistic samples. No real CBCT images were used to train the model; only Gaussian
noise was used. We used a 1,000-step reverse diffusion process to obtain samples, and then we decoded them to create high-resolution 3D volumes.
Results: The CBCT patches created showed realistic anatomical detail, including tooth structures and visible bone
defects. Latent space interpolation showed that the transitions between different types of defects were smooth. The
Fréchet Inception Distance (FID) between the diffusion outputs and the autoencoder reconstructions was 18.4, which
shows that the models were structurally consistent with each other. The average values for the structural similarity index (SSIM) and the PSNR were 0.81 and 28.7 dB, respectively. Training worked well, even with limited GPU resources, and didn't require large datasets.
Conclusion:Our method enables the creation of numerous high-quality synthetic CBCT images without requiring any
clinical data. This simulated framework aids in data augmentation, pretraining, and simulation in dental AI research.
Researchers who work in limited spaces can utilize the model because it is portable and efficient in computer usage. In the future, we will work on conditional generation and validation with expert tasks.
Hi! I can explain how to publish or share, find articles in your field, format citations and answer questions about fees and licensing. How can I help?
Answers are based on The Gufo’s published policies. Editorial decisions are made by the editorial team.
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.
