Deep learning super-resolution for dental CBCT using micro-CT reference and edge loss function

Cone-beam computed tomography (CBCT) is used extensively in dental practice but has limited spatial resolution for visualising fine root canal structures. Micro-computed tomography (micro-CT) offers superior resolution but is unsuitable for clinical use. This study investigated the possibility of enhancing CBCT resolution through deep learning-based super-resolution, using paired micro-CT images as the ground truth.

Two architectures, Enhanced Super-Resolution Generative Adversarial Network (ESRGAN) and Hybrid Attention Transformer (HAT), were trained and evaluated. An edge loss function combining Gaussian and median filtering with Sobel edge detection was introduced to improve structural detail. CBCT and micro-CT images from 48 extracted human teeth were processed into matched datasets. Performance was evaluated using peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), expert visual appraisal, and three-dimensional reconstructions.

Both ESRGAN_edge and HAT_edge significantly outperformed bicubic interpolation and their non-edge-loss counterparts. Subjective ratings indicated that ESRGAN_edge and HAT_edge approached micro-CT quality. Three-dimensional reconstructions confirmed improved anatomical accuracy of pulp chamber and root canal structures, with ESRGAN_edge showing the greatest overlap with micro-CT. Clinical CBCT testing demonstrated that the trained models enhanced root canal clarity, although artefacts in crown regions require further refinement.

Micro-CT-guided super-resolution, particularly with edge optimisation, substantially improved CBCT diagnostic utility in endodontics CLINICAL SIGNIFICANCE: The super-resolution models investigated in the present work achieve acceptable results in enhancing the resolution of the roots of teeth in clinical CBCT scans. Edge-aware super-resolution deep learning models hold promise for clinical dental imaging.