Automatic caries detection in bitewing radiographs-Part II: experimental comparison.

Area of Science:

• Dentistry
• Artificial Intelligence
• Medical Imaging

Background:

• Dental caries detection from bitewing radiographs is crucial for timely intervention.
• Expert consensus on caries detection can be variable, impacting diagnostic accuracy.

Purpose of the Study:

• To compare the diagnostic performance of an automatic deep learning method against multiple dentists for caries detection in bitewing radiographs.
• To evaluate the automatic method's performance in the absence of a definitive ground truth.

Main Methods:

• Four dental experts and three novices manually identified caries using bounding boxes on 100 bitewing radiographs.
• An automatic object detection deep learning model processed the same dataset.
• Annotator performance was evaluated using pairwise comparisons based on error counts and intersection over union (IoU), referencing a consensus standard and the model's training annotator.

Main Results:

• The automatic method outperformed all dentists except the original annotator in terms of mean errors and ranked highly in IoU.
• When compared to a consensus standard, the automatic method showed the best performance in error count and was slightly below average in IoU.
• The automatic method achieved the highest IoU compared to the original annotator, with only one expert making fewer errors.

Conclusions:

• The automatic deep learning method demonstrated consistent superiority over novices and comparable performance to experienced dentists in caries detection.
• Low inter-expert consensus highlights the potential for AI tools to improve the accuracy and repeatability of caries detection, serving as a valuable second opinion.