Does the FARNet neural network algorithm accurately identify Posteroanterior cephalometric landmarks?

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Orthodontics

Background:

• Assessing the accuracy of the Feature Aggregation and Refinement Network (FARNet) algorithm for identifying posteroanterior (PA) cephalometric landmarks.
• Evaluating the potential of AI in automating and improving the precision of landmark identification in cephalometric analysis.

Purpose of the Study:

• To determine the accuracy of the FARNet algorithm in detecting 47 specific anatomical landmarks on PA cephalograms.
• To compare the performance of the FARNet AI algorithm against human expert identification of cephalometric landmarks.

Main Methods:

• Utilized a dataset of 1,431 PA cephalograms, with 1,177 for training, 117 for validation, and 137 for testing.
• Employed a FARNet-based AI algorithm for automatic landmark detection and calculated Mean Radial Error (MRE) and Successful Detection Rates (SDRs) at various thresholds (2-4 mm).
• Conducted statistical analysis, including the Mann-Whitney U test, to compare AI performance with manual identification and analyzed directional differences.

Main Results:

• The FARNet algorithm achieved high SDRs for landmarks like the right gonion (e.g., 100% within 3-4 mm) but lower rates for others, such as the right condylon (e.g., 54.0% within 3 mm).
• AI model accuracy was comparable to human experts, outperforming experts on four skeletal points while experts were superior on one skeletal and seven dental points (P < 0.05).
• Significant deviations were observed along the y-axis for most points, with AI and manual re-identification showing opposite x-axis and same y-axis movements relative to ground truth.

Conclusions:

• The FARNet algorithm effectively identifies key cephalometric landmarks, offering a reliable and potentially more efficient alternative to manual methods.
• The AI system's performance, comparable to human experts, indicates its utility in streamlining orthodontic diagnosis and treatment planning.
• Further research may focus on refining AI performance for specific challenging landmarks and improving directional accuracy.