OPG-based dental age estimation using a data-technical exploration of deep learning techniques

Area of Science:

• Forensic Science
• Computer Science
• Radiology

Background:

• Manual dental age estimation faces challenges like tedium and interobserver variability.
• Automated deep learning methods for age estimation struggle with data scarcity and training complexities.

Purpose of the Study:

• To investigate the impact of convolutional neural network (CNN) hyperparameters on dental age estimation accuracy.
• To evaluate model complexity, batch size, and sample quantity effects on age estimation from orthopantomograms (OPGs).

Main Methods:

• Cross-validation of EfficientNet-B4, DenseNet-201, and MobileNet V3 models on 3896 OPGs.
• Training with batch sizes ranging from 10 to 160 and using random data subsets.
• Analysis of hyperparameter tuning effects on mean absolute error (MAE).

Main Results:

• EfficientNet-B4 achieved the lowest MAE of 0.562 years with a batch size of 160 on the full dataset.
• Increasing batch size improved performance for EfficientNet-B4 and DenseNet-201, while MobileNet V3 peaked at batch size 40.
• Training with complete datasets outperformed reduced sample sizes, highlighting the importance of data quantity.

Conclusions:

• Hyperparameter optimization is crucial for effective deep learning-based dental age estimation.
• Tailored training methodologies and sufficient data are key to achieving accurate forensic age assessment.
• This study advances automated age estimation, demonstrating the potential of optimized CNNs.