Detection of Dental Apical Lesions Using CNNs on Periapical Radiograph

• Main Authors: Li, Chun-Wei, Lin, Szu-Yin, Chou, He-Sheng, Chen, Tsung-Yi, Chen, Yu-An, Liu, Sheng-Yu, Liu, Yu-Lin, Chen, Chiung-An, Huang, Yen-Cheng, Chen, Shih-Lun, Mao, Yi-Cheng, Abu, Patricia Angela R, Chiang, Wei-Yuan, Lo, Wen-Shen
• Format: text
• Published: Archīum Ateneo 2021
• Subjects: biomedical image; periapical image; apical lesion; Gaussian high pass filter; iterative thresholding; deep learning; CNN; Analytical, Diagnostic and Therapeutic Techniques and Equipment; Computer Sciences; Databases and Information Systems; Dentistry; Diseases
• Online Access: https://archium.ateneo.edu/discs-faculty-pubs/278; https://archium.ateneo.edu/cgi/viewcontent.cgi?article=1277&context=discs-faculty-pubs
• Institution: Ateneo De Manila University

Apical lesions, the general term for chronic infectious diseases, are very common dental diseases in modern life, and are caused by various factors. The current prevailing endodontic treatment makes use of X-ray photography taken from patients where the lesion area is marked manually, which is therefore time consuming. Additionally, for some images the significant details might not be recognizable due to the different shooting angles or doses. To make the diagnosis process shorter and efficient, repetitive tasks should be performed automatically to allow the dentists to focus more on the technical and medical diagnosis, such as treatment, tooth cleaning, or medical communication. To realize the automatic diagnosis, this article proposes and establishes a lesion area analysis model based on convolutional neural networks (CNN). For establishing a standardized database for clinical application, the Institutional Review Board (IRB) with application number 202002030B0 has been approved with the database established by dentists who provided the practical clinical data. In this study, the image data is preprocessed by a Gaussian high-pass filter. Then, an iterative thresholding is applied to slice the X-ray image into several individual tooth sample images. The collection of individual tooth images that comprises the image database are used as input into the CNN migration learning model for training. Seventy percent (70%) of the image database is used for training and validating the model while the remaining 30% is used for testing and estimating the accuracy of the model. The practical diagnosis accuracy of the proposed CNN model is 92.5%. The proposed model successfully facilitated the automatic diagnosis of the apical lesion.