Semantic segmentation for tooth cracks using improved DeepLabv3+ model.

Area of Science:

• Biomedical Engineering
• Computer Vision
• Dental Imaging

Background:

• Accurate detection of cracked teeth is crucial for maintaining oral health.
• Existing methods for tooth crack segmentation face challenges in precision and efficiency.

Purpose of the Study:

• To evaluate the performance of an enhanced tooth crack segmentation model, FDB-DeepLabv3+, on optical microscopic images.
• To improve the accuracy and robustness of detecting hidden cracks in teeth.

Main Methods:

• The FDB-DeepLabv3+ model incorporates ResNet50 backbone, Feature Pyramid Network (FPN) for multi-level feature fusion, Densely linked Atrous Spatial Pyramid Pooling (Dense ASPP) for enhanced sampling, and Bottleneck Attention Module (BAM) for local feature extraction.
• The model was tested on a self-made dataset of hidden cracked teeth.

Main Results:

• The FDB-DeepLabv3+ model significantly outperformed four classical networks (FCN, U-Net, SegNet, DeepLabv3+) in segmentation accuracy.
• The proposed network achieved an 11.41% increase in Mean Pixel Accuracy (MPA) and a 12.14% increase in Mean Intersection over Union (MIoU) compared to DeepLabv3+.
• Ablation studies confirmed the beneficial impact of all implemented modifications.

Conclusions:

• The developed FDB-DeepLabv3+ network demonstrates strong performance and robustness for segmenting tooth surface cracks.
• This improved model holds significant potential for advancing computer-aided diagnosis systems for cracked teeth.