Explainable Self-Supervised Learning to Predict and Generate Nutrient Canal Diameters from Mandibular Cone-Beam Computed Tomography Analysis

Introduction: Cone-beam computed tomography (CBCT) has revolutionized dentistry by providing high-resolution 3D views for evaluating mandibular nutrient canal systems. However, manual measurements are time-consuming. This study employs self-supervised learning to predict nutrient canal measurements from mandibular CBCT scans. Objective: To enhance accuracy and treatment planning using explainable artificial intelligence (AI), optimizing strategies for implant surgeries and personalized care. Methods: A total of 398 CBCT images were collected from the DIAS system at Saveetha Dental College. A periodontist annotated data, including nutrient canal diameter, lingual artery visibility, and distances. Data were split into training and testing sets. The self-supervised model utilized autoencoders with encoder, decoder, and regression heads, compressing data into a 16-dimensional latent space to predict canal diameter. Results: The model reduced reconstruction and regression losses, achieving final losses of 0.2543 for reconstruction and 0.3336 for regression. Conclusion: Self-supervised learning can enhance CBCT scan analysis by predicting canal diameters. However, success depends on high-quality data, robust validation, and multimodal integration.