Real-time dental caries segmentation with an efficient Deformable U-Net (DU-Net) for teledentistry system

Abstract

Digital technology has greatly improved teledentistry by facilitating telediagnostics and teleconsultations, particularly benefiting those in remote areas. Additionally, AI advancements enhance diagnostic accuracy and streamline clinical decision-making, reducing costs and resource disparities in dental care. This study presents an improved U-Net architecture, Deformable U-Net (DU-Net), for semantic dental caries segmentation, leveraging deformable convolutions to dynamically adjust sampling points for improved feature extraction and reduced computational redundancy. By connecting encoder-decoder blocks via skip-connections, the DU-Net architecture enables efficient real-time segmentation and balance accuracy while reducing computational demands. The deformable block in DU-Net and DDR U-Net shows a balanced performance and efficiency while maintaining accuracy despite reduced FLOPs. The proposed architecture was implemented in real-time dental caries segmentation on a Dual Core Cortex A72 system and web server. It shows a significant improvement in Dice score, reducing CPU and memory usage compared to conventional U-Net models. Moreover, the DU-Net and its half variants achieved competitive performance with much lower computational demands makes suitable for web servers and embedded applications. The result highlights the DU-Net capability to optimize both computational efficiency and segmentation accuracy, offering a promising solution for real-world applications where speed and resource management are critical, particularly in the medical imaging field.