Automatic Detection of mandibular fractures in panoramic radiographs using a deep convolutional neural network

Abstract

Background and Objectives: Currently, Artificial intelligence and convolutional neural networks (CNNs) have extensive medical applications as in detection and diagnosis of diseases and clinical disorders. The present study aimed to detect mandibular fractures on panoramic radiographs using CNN. Materials and Methods: This study evaluates 275 panoramic radiographs retrieved from the archives of the Oral and Maxillofacial Radiology Department of Tabriz School of Dentistry. From all of the radiographs, 124 have mandibular fractures and 151 have no fracture. First, the location of mandibular fractures is detected and annotated on the radiographs by oral and maxillofacial radiology resident, then all of the annotated images is re-examined by the oral and maxillofacial radiologist. Next, noise reduction was performed using the Chebyshev type II filter. To standardize the images, their primary resolution is modified and converted to 227 x 227. The 32 layer AlexNet CNN is then used for training and primary classification of images with and without mandibular fractures. 53 layer Alexnet CNNis subsequently used to detect the location of mandibular fractures on images. Of all images, 60% are randomly used for network training, 20% for validation, and 20% for final testing. The precision, recall, and F1 score are measured to assess the efficacy of this algorithm for detection of mandibular fractures. Results: The precision, recall, and F1 score of the algorithm for detection of mandibular farctures is found to be 0.968, 0.834, and 0.896, respectively. Conclusion: The suggested algorithm successfully detectsmandibular fractures on panoramic radiographs with high accuracy. So Models based on CNNs are expected to enhance the detection of mandibular fractures on panoramic radiographs.