Deep Learning-Driven Super-Resolution for Dental Cone-Beam Computed Tomography: An Ex-vivo Proof-of-Concept Study using Artificially Degraded Micro-Computed Tomography Data

Objectives The current proof-of-concept study aims to develop deep learning (DL)-based super-resolution (SR) models to enhance simulated cone-beam computed tomography (CBCT) images, derived from degraded micro-computed tomography (micro-CT). Methods For this ex vivo study, we collected micro-CT data of 51 extracted teeth and then artificially degraded them using blurring and downscaling to simulate CBCT images. Three DL models, Super-Resolution Convolutional Neural Network (SRCNN), Local Texture Estimator (LTE), and Swin Transformer for Image Restoration (SwinIR), were trained and compared with bicubic interpolation. Image quality was assessed using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). Moreover, three dentists evaluated the images’ sharpness and noise using a 5-point Likert scale. Results All DL models significantly outperformed bicubic interpolation and low-resolution images in objective metrics. SwinIR showed superior PSNR (30.36 ± 2.66), closely followed by LTE (30.34 ± 2.69), while SRCNN achieved the highest SSIM (0.889 ± 0.073). Subjectively, LTE and SwinIR both scored a median sharpness of 4, with LTE excelling in the mean score (3.79 ± 0.47). For noise reduction, SRCNN performed best among the DL models (median = 4), still lower than bicubic interpolation (median = 5). Conclusion This study demonstrated that DL-based SR models can effectively enhance simulated dental CBCT images towards micro-CT quality. Among the selected models, LTE demonstrated superior perceptual sharpness. Although validation on clinical CBCT images and diagnostic tasks remains necessary, these findings establish technical feasibility and provide a foundation for future translational studies.