International Conference on Information and Knowledge Technology

Home / شانزدهمین کنفرانس بین المللی فناوری اطلاعات و دانش

A U-Net architecture with graph attention networks to accurately define tooth boundaries

Authors :

Ehsan Akefi¹ Hassan Khotanlou²

1- دانشگاه بوعلی سینا همدان 2- دانشگاه بوعلی سینا همدان

Keywords :

Image Segmentation،U-Net،Graph Neural Network،Graph Attention Network،Dental Panoramic Radiography

Abstract :

It is very important for clinical diagnosis and treatment planning to be able to accurately segment teeth in panoramic radiographs. However, this is still a big problem because teeth often overlap, and standard convolutional neural networks (CNNs) have trouble capturing long-range spatial dependencies. This paper presents a novel hybrid deep learning architecture that combines a U-Net with an Advanced Spatial Graph Processor to address these constraints. The proposed model substitutes the conventional bottleneck of the U-Net with a Graph Neural Network (GNN) module, which distinctly represents non-local relationships among various regions of the image by converting the feature map into a graph structure. The model can dynamically focus on important structural patterns by using Graph Attention Networks (GAT). This makes it much easier to see the boundaries of complex and overlapping teeth. To address the issue of insufficient labeled medical data, a comprehensive data augmentation pipeline was implemented. This increased the training dataset by five times, making the model more generalizable. Our hybrid approach is better than the other one, as shown by experimental results on the Tufts Dental Database. The proposed model with attention (Unet + graph + attention) outperformed the baseline U-Net, achieving a Dice Score of 92.91% and an Intersection over Union (IOU) of 86.77%. These results show that using the local feature extraction capabilities of U-Net with the global structural modeling of GNNs is a strong and very accurate way to segment teeth. This has a lot of potential for use in clinical settings.