Enhancing Yolov8 Backbone Using Gradient Descent-based Method for Dental Segmentation

This research addresses the challenge of dental segmentation in computer vision, a task focused on accurately outlining dental structures in images. The traditional methods, particularly convolution neural networks (CNNs), often suffer from suboptimal performance and computational inefficiency. Our study introduces an enhanced approach by applying the YOLOv8 algorithm, known for its effectiveness in object detection, for dental segmentation. Our proposed model improves YOLOv8’s feature extraction capability by integrating additional layers into its backbone architecture, primarily focusing on the Coordinates-To-Features (C2f) module. This C2f-based feature extraction technique is designed to optimize gradient descent, reducing loss and maximizing prediction accuracy. By incorporating adaptive weights, the model effectively enhances the propagation of gradients, allowing for a more precise focus on dental structures. The adapted model, comprising 29 layers, is trained on a large-scale real-color dental dataset. Experimental evaluation demonstrates that the proposed model achieves exceptional performance, attaining 99.6% precision and recall in dental segmentation tasks. These results highlight the potential of YOLOv8 for specialized segmentation challenges and mark a significant contribution to automated dental analysis, offering direct benefits for clinical diagnostics and treatment planning in dentistry.