In dentistry, panoramic radiography or orthopantomography (OPGs) is an essential adjunct to first-line dental screening and dental treatment planning. It offers immediate availability, limited radiation dose, and the possibility of enhancement when post-processed, making it the preferred diagnostic tool for dental professionals [1,2].

Nevertheless, even if these images undoubtedly aid practitioners, they also have certain limitations that can hinder the decision-making process, such as bidimensionality, potential artifacts, or lack of homogeneity in regions of interest. Therefore, the development of automated assistance models could help reduce both intra- and inter-rater variability, allowing consistency and a more reliable and accurate diagnosis, especially for inexperienced professionals [3].

Recently, artificial intelligence (AI) has demonstrated its ability to achieve significant positive outcomes in healthcare, notably in dentistry. This might be especially interesting when considering the potential for misdiagnoses arising from inexperience or general fatigue [4,5]. AI can be defined as the capacity of machines to imitate both human knowledge and decision-making processes to achieve certain outcomes, although an accurate definition for these systems has been reported to be somewhat challenging to describe [5]. Overall, this technology can be classified into machine learning (ML) and deep learning (DL). Because ML requires previous manual feature extraction, using it for automatic radiograph interpretation is time-consuming and expensive [2,[6], [7], [8]].

AI is gradually being implemented in nearly every dental specialty, and its application has already been studied for bone-loss detection [9], dental implant classification [7], taurodont teeth detection [10], and periapical pathology identification in both 2D and 3D radiological imagery [11]. For this purpose, neural networks have to be trained to automatically identify dental structures and treatments [12], as well as the main maxillofacial anatomical structures [2,4,5,13], which can be considered one of the most challenging steps in the development of AI-based systems capable of interpreting images. Thus, this should be performed as accurately as possible, with and without ideal analytical conditions [14].

Despite the increasing interest of researchers in this topic, there remains a huge gap in information concerning the different back-end neural network architectures currently being used and implemented to achieve their objectives. In addition to the interesting results obtained by AI, little is known about CNNs that could obtain the best results with a reduced dataset when performing multiclass detection and segmentation tasks.

Faster region CNN (Faster R-CNN) [15], ResNet [1], and U-Net [16] architectures are three of the most widely used neural networks. Nevertheless, OPG object detection and segmentation tasks have only been reported to be used synergistically altogether [1] in the Visual Geometry Group [17] and You Only Look Once (YOLO) architecture families [4]. The last neural network is a real-time object detection algorithm that can simultaneously detect multiple objects in an image without requiring any reference datasets as input because it can be freely trained directly from the validation data. It has been widely used for this matter owing to its high detection speed, accuracy, and generalization ability [18,19]. Its seventh version (YOLOv5) was released in 2020 and has been updated periodically until its latest form, YOLO7, which has been reported to achieve excellent performance results compared to the previously mentioned version [18]. To date, this is the only neural network that has been used individually for OPG multiclass object detection and labeling using datasets of over 1000 images [19]. To the best of our knowledge, no previous study has used this architecture to describe whether its results for segmentation purposes are as positive as those for object detection, while using both large and reduced datasets.

Therefore, the present study aimed to report the results of diagnostic performance evaluation of the three latest versions of YOLO convolutional neural network in terms of object detection and segmentation when utilizing a multiclass and reduced panoramic radiograph dataset. Thus, the null hypothesis was established as there were no significant differences in diagnostic performance between the included YOLO versions.