Evaluation of artificial intelligence for detecting periapical lesions on panoramic radiographs

Sažetak

Background/Aim: The aim of this study is to evaluate the function of diagnostic computer software designed for the detection of periapical lesions on panoramic images with deep learning methods. Material and Methods: In our study, more than 500 adult retrospective panoramic radiography images obtained randomly were evaluated, and periapical lesions were labeled on the radiographs with the ResultLab.Ai labeling program (ResultLab.AI, Istanbul, Turkey). All images have been rechecked and verified by Oral, Dental and Maxillofacial Radiology experts. The model used is a U-Net based architecture customized for dental radiographs and optimized for fast operation. What is meant by customization is the structures called "Attention Gate" added to the architecture to draw the model's attention to the lesions. Results: Mask estimation was made separately for each image and F1 and IoU scores were calculated by comparing them with the marked masks. A list was created from the calculated F1 and IoU scores. The F1 and IoU score of the entire data set was calculated by taking the average of the values in this list. In IoU score calculation, Keras library's Mean IoU metric was used to calculate for 2 classes. In the F1 score calculation, the F1 score metric of the SKLearn library was used to calculate the unweighted average for 2 classes. As a result, the IoU-Score was obtained as 0.8578 and the F1-Score as 0.8587. Conclusions: Detection of periapical lesions on panoramic radiographs with an artificial intelligence approach will help practitioners reach definitive conclusions even in lesions that are likely to be overlooked. In addition, success rates for deep learning methods improve as data set size grows. As the number of images rises, the training models' success rates will follow.