Effect of Lower Third Molar Segmentations on Automated Tooth Development Staging using a Convolutional Neural...

Area of Science:

• Forensic Odontology
• Radiology
• Artificial Intelligence in Medicine

Background:

• Third molar development staging is crucial for age estimation in subadults.
• Previous automated methods using bounding boxes (AlexNet CNN) included surrounding structures, potentially limiting accuracy.
• Hypothesis: Segmenting only the third molar can improve automated staging performance.

Purpose of the Study:

• To determine and validate the effect of lower third molar segmentation on automated tooth development staging.
• To compare the performance of different segmentation methods (bounding box, rough, full) and CNN architectures (AlexNet, DenseNet201).

Main Methods:

• Retrospective collection and processing of 400 panoramic radiographs.
• Segmentation of lower third molars using three methods: bounding box (BB), rough segmentation (RS), and full segmentation (FS).
• Automated stage allocation using DenseNet201 Convolutional Neural Network (CNN), with results compared to human-allocated reference stages.

Main Results:

• Full tooth segmentation (FS) achieved the highest accuracy (0.61), lowest mean absolute difference (0.53 stages), and highest Cohen's linear kappa (0.84).
• DenseNet201 improved accuracy by 3% over AlexNet (BA vs. BB).
• FS increased correct stage allocation by 7% compared to BB, with most misallocations being neighboring stages.

Conclusions:

• Full tooth segmentation significantly optimizes automated dental stage allocation for age estimation.
• DenseNet CNN architecture further enhances the performance of automated staging.
• The findings support the use of refined segmentation and advanced CNNs for more accurate forensic age assessment.