Dental bitewing radiographs segmentation using deep learning-based convolutional neural network algorithms.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Dental image segmentation is crucial for diagnosis but challenging due to image quality issues.
• Deep learning models offer promising advancements in analyzing complex dental images.

Purpose of the Study:

• To evaluate and identify the most effective deep learning segmentation technique for bitewing radiographs.
• To compare segmentation performance based on accuracy, training time, and model complexity.

Main Methods:

• Employed deep learning models: Resnet-18, Resnet-50, Xception, Inception Resnet v2, and Mobilenetv2.
• Utilized MATLAB for image preprocessing and graph cut segmentation to create binary masks.
• Trained and validated models on 298 and 99 radiographs, respectively, with testing on 99 images.

Main Results:

• Resnet-18 and Resnet-50 models demonstrated high segmentation accuracy at 93.67% and 94.42%, respectively.
• Performance was evaluated based on accuracy, speed, and model size (number of parameters).
• Findings were compared with previous research to highlight advancements in dental image segmentation.

Conclusions:

• Resnet-50 and Resnet-18 are highly effective for segmenting bitewing radiographs, offering superior accuracy.
• This research provides guidance for selecting optimal segmentation methods in practical dental image analysis.
• The study contributes to advancing dental diagnostics and treatment planning through improved image analysis.