EGT-UNet: Evolutionary Game-Theoretic Adaptive Optimization for Pediatric Panoramic Tooth Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate segmentation of pediatric panoramic teeth is crucial for dental diagnosis and treatment planning.
• Existing methods face challenges due to the limited availability of high-quality pediatric dental datasets and the complexity of mixed dentition.

Purpose of the Study:

• To propose EGT-UNet, an innovative deep learning model for pediatric panoramic tooth segmentation.
• To introduce a novel pediatric panoramic image database to address data scarcity.
• To enhance segmentation accuracy through a dynamic loss optimization mechanism based on Evolutionary Game Theory (EGT).

Main Methods:

• Development of a new pediatric panoramic image database with 1269 pixel-level annotated images.
• Implementation of EGT-UNet, an enhanced U-Net architecture incorporating Squeeze-and-Excitation blocks, Attention Gates, and dilated convolutions.
• Dynamic tuning of loss weights (Dice, Focal Tversky, Boundary) using replicator dynamics and task-specific fitness functions.

Main Results:

• The proposed EGT-UNet achieved high performance, with the best model ('EGT Aggressive') reaching Dice = 0.931 ± 0.044 and IoU = 0.873 ± 0.066.
• A statistically significant superiority was observed over baseline networks in region and boundary metrics (Wilcoxon p < 0.05).
• The model demonstrated increased robustness in segmenting complex mixed dentition images.

Conclusions:

• EGT-UNet offers a significant advancement in pediatric panoramic tooth segmentation.
• The dynamic loss weighting strategy based on EGT improves model performance and robustness.
• The newly created dataset and proposed model contribute valuable resources to the field of pediatric dental imaging.