High-Fidelity Synthesis of Temporomandibular Joint Cone-Beam Computed Tomography Images via Latent Diffusion Models

Background: The development of robust artificial intelligence (AI) models for diagnosing Temporomandibular Disorders (TMDs) is severely constrained by data scarcity and patient privacy regulations. Cone-beam computed tomography (CBCT), the gold standard for assessing osseous changes in the temporomandibular joint (TMJ), inherently contains sensitive biometric facial features, making de-identification difficult without losing critical anatomical information. This study aims to develop and evaluate TMJCTGenerator, a specialized latent diffusion model (LDM) framework designed to synthesize high-fidelity, diverse, and anonymous TMJ CBCT images. We hypothesize that this LDM approach can achieve superior anatomical fidelity and diversity compared to traditional generative adversarial network (GAN)- and variational autoencoder (VAE)-based methods, specifically in capturing fine osseous details within sagittal and coronal views of the mandibular condyle. Methods: A training dataset comprising 348 anonymized CBCT volumes was obtained in this retrospective comparative study to extract high-resolution sagittal and coronal regions of interest of the mandibular condyle. An independent test set of 39 anonymized CBCT volumes was further included. We developed a class-conditional LDM that integrates a pre-trained VAE for perceptual compression with a conditional U-Net for iterative denoising in the latent space. Performance was evaluated via qualitative anatomical fidelity assessment, Fréchet Inception Distance (FID), and a blinded Visual Turing test conducted by experienced clinicians to determine the distinguishability of synthetic images from real data. Results: Qualitative analysis revealed that TMJCTGenerator produced images with superior sharpness and anatomical consistency compared to baseline models, successfully reconstructing fine bone structures essential for diagnosing degenerative joint disease. TMJCTGenerator achieved lower FID scores than both VAE and GAN baselines. In the visual Turing test, clinicians were unable to reliably distinguish the generated images from real scans, and non-inferiority analysis confirmed that the synthetic data were statistically non-inferior to real data. Furthermore, TMJCTGenerator demonstrated the capability to generate diverse pathological conditions, ranging from normal anatomy to severe osteoarthritic changes. Conclusions: The proposed LDM framework effectively addresses the data scarcity and privacy bottlenecks in TMJ AI research by generating realistic, fully anonymous medical imaging data. TMJCTGenerator outperforms traditional generative methods in both visual fidelity and diversity, offering a viable solution for training downstream diagnostic algorithms. The source code and pre-trained models of TMJCTGenerator have been made open-source.