There are some problems with implants that cannot be solved in general clinics. In these problems, an unknown implant system will make the problems worse. Therefore, the identification of an implant system is necessary for both dentists and patients, and an automated identification system that is not dependent on the dentist’s expertise is needed. Considering these issues, an AI-based approach seems to be a potentially suitable solution, and this study was conducted to focus on developing an automated identification system of implants from panoramic radiographic images using object detection. There are already two methods for implant identification [14, 15]. In the first, dental radiographic images of many implant systems have been uploaded to a website and dentists are able to check them to find an image that matches the patient’s implant image. The second system employs nine questions about implant characteristics. The database returns candidate matching implants based on the answers to these questions, and dentists must match them with those of the patient. Both of these systems require dentists to check whether two images of an implant are the same to identify the implant system. In contrast, the system in this study is based on deep learning, one of AI techniques, and not a dentist but the computer itself identifies the implant.
When evaluating the performance of the object detection, two indices, mean average precision (mAP) and mean intersection over union (mIoU), were mainly used. mAP is used to measure the accuracy of object detection model, and the closer the value is to 1.0, the more accurate the model is. A mAP of more than 0.7 seemed to be regarded as a good value in other studies [13], but there is no clear criterion. A mIoU of more than 0.7 is regarded as a good value [16, 17], and the mAP the mIoU obtained in this study are 0.71 and 0.72, respectively. Considering these, the performance of this learning system can be considered to be high. The values of the hyperparameters were determined from the results of preliminary experiments with various combinations of values. Learning with this combination yielded superior mIoU and mAP values than other combinations.
In the results of this study, the AP of each implant system varies from 0.51 for Genesio to 0.85 for MK III/IIIG, and the mAP is 0.71. The TP ratios also vary from 0.50 for Genesio to 0.82 for MK III/IIIG. These differences are caused by the number of implants, their locations, and their similarity of shape. When selecting implant systems to recognize in this study, frequently used implant systems were selected because the number of implants seemed to be one of the most important factors. In fact, both AP and TP ratio of Genesio, which was the least number of images, were the minimum value, and those of MK III/IIIG were the maximum. About 1300 panoramic images and a total of 3000 implant images were used, but these numbers were not enough to recognize all the implant systems included. To increase the learning performance, a sufficient number of implant images are necessary.
To identify implant systems from radiographic images, dental radiography, panoramic radiography, and computed tomography were considered. In this system, it is thought that implant systems are identified by the shape of the collar, groove, and apex of the implant images, which are unique characteristics of each implant. Consequently, the quality of the training images is important so that these shapes of the implants can be recognized in detail. The advantage of using panoramic radiographic images is that they are standardized to a certain level regardless of the patient, and the shapes of the implants in the images are also standardized. However, the disadvantage is that the implant shapes are unclear when they overlap with a shadow, such as the spina or floor of the maxillary sinus, or when they were too short or much inclined. This may cause misdetection, and some misdetections actually occurred in the result of this study (Fig. 6). In such cases, the images of dental radiography may be more useful. Another disadvantage is the shape of the images. The shape of the images in the learning procedure of this algorithm is square, but the original panoramic radiographic images are rectangular. Therefore, in the learning procedure, panoramic radiographic images are laterally compressed and the shapes of the implants are also compressed. As a result, implant details become unclear, and this could decrease learning performance. The learning performance could be increased by cropping the original panoramic image into a square shape that includes implants beforehand.
In this study, four systems by one manufacturer and more two systems by two manufactures were selected. The reason was to know how much of a difference this system could identify. In the results, the misidentification between MK III/IIIG and Genesio often occurred, especially some of Genesio were misidentified as MK III/IIIG. They are all straight type, and the differences among them are subtle: differences among three systems are the shape of the platform and apex. These small differences are not easy to distinguish in compressed images and misidentification hence occurred. Increasing images with high quality must also be a solution to prevent these misidentifications. In addition to the shape of apex or collar, other differences, such as the shape of the inner screw or space between the bottom of the inner screw and implant body, may be helpful to identify similar-shaped implant.