Automatic detection of degenerative changes in the temporomandibular joint region using deep learning with panoramic radiographs

Background/Aim: The temporomandibular joint (TMJ) is a complex anatomical region composed of the mandibular condyle located in the glenoid fossa of the temporal bone and covered with fibrous connective tissue. Excessive and continuous forces lead to progressive degeneration of the bony surfaces of t...

Full description

Saved in:

Bibliographic Details
Published in:	Balkan Journal of Dental Medicine Vol. 28; no. 2; pp. 99 - 116
Main Authors:	Tassoker, Melek, Hakli, Huseyin, Yaman, Metin, Ekmekcı, Sema, Incekara, Senanur, Kamaci, Serhat, Ozturk, Busra
Format:	Journal Article
Language:	English
Published:	Balkan Stomatological Society 2024
Subjects:	deep learning degeneration panoramic radiography temporomandibular joint
Online Access:	Get full text
Tags:	Add Tag No Tags, Be the first to tag this record!

Description
Summary:	Background/Aim: The temporomandibular joint (TMJ) is a complex anatomical region composed of the mandibular condyle located in the glenoid fossa of the temporal bone and covered with fibrous connective tissue. Excessive and continuous forces lead to progressive degeneration of the bony surfaces of the TMJ. The aim of this study is to determine the success of automatic detection of degenerative changes detected on panoramic radiographs in the TMJ region with deep learning method. Material and Methods: Panoramic images of 1068 patients (1000 with normal TMJ appearance and 68 with TMJ degeneration) over 18 years of age were included in the study. CVAT, open-source annotation tool (https://www.cvat.ai/) was used for labeling image data. All images were resized using the bilinear interpolation method. With the using data augmentation techniques, the number of images data reached 1480. BSRGAN model was applied to the data to increase the resolution of the data. YOLOv5, YOLOv7 and YOLOv8 algorithms were used for TMJ degeneration detection. TP, FP, TN, FN, accuracy, precision, recall, F1-score and AUC (Area Under the Curve) metrics were used for statistical analysis. Results: YOLOv5s training resulted in 94.40% accuracy, 81.63% precision, 86.96% sensitivity, 84.21% F1 score and 91.45% AUC. YOLOv7 training resulted in 99.63% accuracy, 97.87% precision, 100% sensitivity, 98.92% F1 Score and 99.77% AUC. YOLOv8 training resulted 96.64% accuracy, 91.11% precision, 89.13% sensitivity, 90.11% F1 Score and 93.66% AUC. Conclusions: All three algorithms have high success rates, with the best results obtained in YOLOv7.
ISSN:	2335-0245 2738-0807 2335-0245
DOI:	10.5937/bjdm2402099T