Evaluating Deep Learning AI for Periapical Lesion Detection Across Panoramic, Periapical, and CBCT Radiographs: A Systematic Review

1. INTRODUCTION

Artificial Intelligence (AI) refers to the capability of computational systems to emulate human cognitive functions in performing specialized tasks. Within the healthcare domain, AI technologies have advanced significantly and are increasingly being implemented across medical and dental disciplines. In dentistry specifically, AI applications are frequently deployed through Deep Learning (DL) methodologies, a subset of machine learning that has demonstrated particular efficacy in dental diagnostics and image analysis [1].

Deep Learning (DL), a specialized subset of Artificial Intelligence (AI), has demonstrated significant utility in diagnostic applications, particularly within dentomaxillofacial radiology. Current evidence indicates its substantial potential to enhance both accuracy and efficiency in detecting periapical lesions across multiple radiographic modalities. The fundamental advantage of DL lies in its capacity for autonomous feature extraction from raw data, effectively emulating human cognitive processes without requiring explicit programming. Through this mechanism, AI systems can systematically analyze datasets to perform critical functions, including image classification, enhancement, and segmentation. Notably, DL's sophisticated feature extraction capabilities serve to reduce data dimensionality, thereby transforming complex patterns into quantifiable parameters that can be effectively processed by machine learning algorithms [1].

Among various Deep Learning (DL) architectures, Convolutional Neural Networks (CNNs) represent a specialized class of artificial neural networks particularly suited for image analysis tasks. In medical imaging applications, CNNs have demonstrated exceptional performance in fundamental operations including lesion detection, tissue segmentation, and pathological classification [2]. Regarding detection capabilities, CNN algorithms precisely localize target structures in radiographic images by generating bounding boxes and assigning appropriate diagnostic labels to identified objects.

Periapical lesions are pathological changes that occur in the tissues surrounding the apex of a tooth, most commonly as a consequence of pulp necrosis or bacterial infection. These lesions are frequently identified radiographically as areas of radiolucency. İçöz et al. described periapical inflammatory lesions as a local response in the bone around the root apex due to pulp necrosis [3]. Early and accurate detection of such lesions is essential for effective treatment planning, as delayed diagnosis can lead to disease progression and adverse clinical outcomes. Bayrakdar et al. emphasized the diagnostic importance of periapical radiolucency, noting that it is a key feature of apical periodontitis [4].

Panoramic radiography offers several advantages, including comprehensive visualization of maxillofacial structures, a simplified acquisition technique, applicability for patients with trismus or limited oral opening, and a reduced radiation dose compared to CBCT. However, this modality presents notable limitations: 1)inferior spatial resolution compared with periapical radiography, 2) non-uniform magnification across the image, compromising measurement reliability, and 3) higher radiation exposure than periapical radiography.

Cone-Beam Computed Tomography (CBCT) offers significant diagnostic advantages, such as: 1) distortion-free imaging with multiplanar reconstruction capability (axial, sagittal, coronal, and 3D views), 2) minimal artifact production [5], and 3) high-resolution three-dimensional visualization of anatomical structures with lower radiation exposure than conventional CT. However, CBCT presents notable limitations when compared to 2D radiography, including increased radiation dose, higher operational costs, and prolonged interpretation time due to the substantial volume of three-dimensional anatomical data requiring comprehensive analysis. The extensive datasets generated by CBCT, while diagnostically valuable, require more meticulous evaluation compared to conventional two-dimensional imaging modalities.

While panoramic, periapical, and Cone-Beam Computed Tomography (CBCT) are established modalities for assessing dental structures and surrounding tissues, current evidence lacks consensus on the optimal imaging modality for Artificial Intelligence (AI)-assisted radiodiagnosis of periapical lesions. Each modality presents distinct advantages and limitations that may differentially influence machine learning model performance, necessitating further comparative investigation to determine their respective efficacies in AI applications.

As such, this systematic review aims to comprehensively evaluate and synthesize the current evidence on the application of deep learning models for periapical lesion detection across three radiographic modalities: periapical, panoramic, and Cone-Beam Computed Tomography (CBCT). Through rigorous analysis of the available scientific literature, this study seeks to establish a standardized framework for comparing the diagnostic performance of these imaging techniques when integrated with artificial intelligence algorithms.

Accordingly, a critical understanding of each radiographic modality's comparative advantages and limitations is essential for evaluating how Deep Learning (DL) models might enhance the diagnostic sensitivity and specificity of periapical lesion detection. This systematic review synthesizes current evidence from the scientific literature to: 1) establish standardized performance benchmarks across imaging modalities, 2) identify optimal DL architectures for different radiographic techniques, and 3) delineate future research directions. The findings are anticipated to serve as a foundational reference for both dental researchers developing next-generation DL algorithms and clinicians seeking to implement AI-assisted diagnostic systems in endodontic practice.

2. MATERIALS AND METHODS

This systematic review, conducted from March to May 2024 following PRISMA guidelines, employed the PICO framework to evaluate radiographic modalities for periapical lesion detection. The population comprised studies on periapical lesions, with interventions analyzing CBCT, periapical, and panoramic radiography. Comparisons assessed relative efficacy in early detection, while outcomes determined the optimal modality for AI-assisted diagnosis using deep learning models, particularly CNNs. Electronic database searches systematically identified literature examining AI integration with these modalities to compare diagnostic performance in lesion identification.

The literature search was systematically conducted across three major databases (PubMed, Scopus, and EBSCO) using the following keyword combinations: 1) “periapical lesion” AND “CBCT”, 2) “periapical lesion” AND “periapical radiography”, and 3) “periapical lesion” AND “panoramic radiography”. This search strategy yielded an initial pool of 20 relevant articles, which were subsequently subjected to a rigorous systematic review according to the predefined protocol.

The comprehensive database search was executed in April 2024 across PubMed, Scopus, and EBSCO using predefined keyword combinations. An initial filter limited results to publications from 2019 to 2024, yielding 121 articles. After applying the inclusion criteria and removing duplicates, 72 articles remained. Following rigorous title/abstract screening, 20 studies met all eligibility criteria for final inclusion as shown in Table 1 below.

This systematic review protocol was registered in the International Prospective Register of Systematic Reviews (PROSPERO) under the registration number: CRD420251112376.

3. RESULTS

The final analysis included 20 cross-sectional studies conducted across seven countries (Turkey, the United Kingdom, the United States, South Korea, the Netherlands, Germany, and Switzerland). All selected publications were Scopus-indexed journals (Q1-Q3 tiers), published between 2019 and 2024 in English. These studies are systematically presented in Table 3 [3-23], which provides a comprehensive overview of their key characteristics and findings.

The systematic review identified 20 eligible articles, which are comprehensively summarized in Table 4. The analysis revealed that 17 studies [1-10, 12, 14-15, 17-20] employed Convolutional Neural Networks (CNNs), while the remaining 3 studies [11, 13, 16] utilized U-net architectures for periapical lesion detection.

4. DISCUSSION

Periapical radiography remains one of the most commonly used imaging modalities in dental practice. It offers high-resolution visualization of dental structures, fast image acquisition for immediate clinical interpretation, relatively low radiation exposure, and is cost-effective compared to more advanced modalities like CBCT. According to Song et al., panoramic radiographs are commonly used to diagnose apical lesions and have relatively low radiation doses compared to CBCT. These advantages support its continued use as a frontline diagnostic tool [7].

However, the two-dimensional nature of periapical radiographs imposes inherent limitations. This modality does not provide volumetric anatomical detail, which can lead to missed diagnoses, particularly when anatomical structures overlap. Bayrakdar et al. pointed out that panoramic radiographs, being two-dimensional, may fail to detect certain pathologies due to anatomical superimposition [4]. Moreover, periapical radiographs are highly technique-sensitive; improper angulation or patient positioning may cause geometric distortion. İçöz et al. further noted that these techniques are limited by the nature of two-dimensional imaging for three-dimensional structures, compounded by factors such as anatomic noise, superimposition, and distortion effects [3].

Considering both the advantages and limitations, periapical radiography remains highly valuable for initial diagnostic screening, while CBCT should be reserved for complex cases that require comprehensive three-dimensional anatomical assessment.

In the analysis of periapical radiographs using a deep learning approach on a dataset of 1,169 samples, Fu, W. T. et al. [2024] reported a sensitivity of 87–90% for detecting dental caries, crowns, and dentin pulp; 85–92% for identifying periapical lesions; and 96–98% for detecting root canal fillings17. In the study by Chen, H. et al [2021], improved results were observed in the detection of advanced periapical periodontitis 12. Meanwhile, Hamdan, M. H. et al [2022], in a study involving 68 periapical radiographs validated with Cone Beam Computed Tomography, found that sensitivity varied according to lesion size, 85% for small lesions measuring between 2 and 5 millimeters, and 92% for larger lesions measuring 5 millimeters or more [22]. Moidu, N. P. et al. [2022], using 540 periapical radiographs, reported a sensitivity of 92.1%, a specificity of 86.4%, and an F1 score of 0.86 [10]. Similarly, Pauwels, R. et al., [2021], in an analysis of 5,000 periapical radiographs, demonstrated the ability to detect lesions and estimate their size within 1 minute 45 seconds to 3 minutes, achieving a sensitivity between 79% and 100% and a specificity between 88% and 100% [13].

Elaborating further, Ver Berne et al. [2023] implemented deep learning on 255 panoramic radiographs [197 normal and 58 with radiolucent lesions], achieving 95% sensitivity and specificity for detecting radicular cysts, and 77% sensitivity with 100% specificity for granulomas [9]. Lee, J. H. et al. evaluated 34 patients with odontogenic cysts and keratocysts using both panoramic and CBCT imaging, reporting CBCT sensitivity of 96.1% and specificity of 77.2%, compared with 88.2% sensitivity and 77% specificity with panoramic images [15]. Song, I. S. et al., using 180 panoramic radiographs, identified 147 lesions with an F1 score of 0.82 [7]. Bayrakdar, I. S. et al. analyzed 470 radiographs and reported 92% sensitivity, 84% precision, and an F1 score of 0.8 4. Orhan Kan et al. examined 4,497 teeth with CBCT to assess various dental conditions, achieving 80% sensitivity for detecting healthy and treated teeth, though performance declined for pathological cases [8]. İçöz, D. et al. evaluated 306 panoramic images [400 roots] with apical periodontitis, achieving an F1 score of 0.71. Their AI model detected 119 mandibular and 56 maxillary lesions, with a recall of 0.98, precision of 0.56, and F-measure of 0.71 [3].

In a deep learning CNN-based study by Fu, W. T. et al. [2024], CBCT examinations involving 4,951 teeth [563 with apical periodontitis and 4,388 without] achieved a sensitivity of 88–94% [17]. The CNN model significantly reduced diagnostic time by 69.9 minutes for junior endodontists and 32.4 minutes for senior endodontists. Meanwhile, Wang, X. et al., using 90 CBCT scans, reported 98% accuracy in detecting apical periodontitis [38%], periapical infections [17%], and dental pulp conditions [29%] [23]. Ezhov, M. et al., analyzing 1,346 CBCT images, achieved 85% sensitivity and 96% specificity in identifying periapical lesions [20]. Further, Calazans, M.A. et al., using 5,343 CBCT scans, categorized 454 as normal, 276 as small lesions [0.5–1.9 mm], and 270 as large lesions [≥3 mm] [21]. Setzer, F.C. et al., evaluating 61 roots, reported 93% accuracy and 88% specificity [19]. Huang, J. et al. compared three deep learning models, such as BALD, ME, and Bayesian U-Net, on 100 CBCT slices. BALD exhibited the highest sensitivity [84%], ME achieved the highest specificity [96.3%] and precision [95%] [16]. Hadzic et al. [2023] reported 86.7% sensitivity and 84.3% specificity for CBCT-based detection of periapical lesions [18].

The diagnostic capabilities of junior dentists are significantly enhanced through the integration of artificial intelligence and deep learning techniques applied to CBCT imaging, thereby markedly reducing diagnostic time. Moreover, these methods produce clinically relevant segmentation outcomes [17]. Convolutional Neural Network (CNN) algorithms in CBCT image segmentation have been shown to improve the accuracy, efficiency, and overall quality of oral lesion segmentation compared to conventional approaches such as manual and threshold-based segmentation [23]. Deep learning (DL) applications in periapical lesion detection on CBCT images also demonstrate the potential to enhance diagnostic precision and segmentation quality [19]. In a study by Calazan et al., the combination of CNN with a Siamese Concatenated Network for classifying the presence or absence of periapical lesions in CBCT images achieved an accuracy of 70% [21]. This combined model shows promise as an effective automated classification approach and could support clinicians in achieving faster, more accurate diagnoses of periapical lesions in clinical practice.

In sum, integrating artificial intelligence with panoramic, periapical, and Cone Beam Computed Tomography (CBCT) imaging modalities significantly enhances diagnostic accuracy and operational efficiency in identifying periapical lesions. By applying machine learning algorithms trained on heterogeneous, comprehensive datasets, these systems can assist clinicians in detecting lesions as small as 1 millimeter and distinguishing between benign and malignant pathological conditions. This technological advancement offers several notable benefits, including expedited diagnostic processes, reduced likelihood of human diagnostic errors, and improved precision in identifying cystic lesions. Nevertheless, it is imperative to underscore that such systems are intended to function as supportive tools within clinical practice rather than as replacements for clinical judgment. Accordingly, the responsibility for diagnosis and therapeutic decision-making must reside with qualified dental professionals.

CONCLUSION

In conclusion, artificial intelligence holds the potential to transform the diagnosis of periapical lesions by augmenting the diagnostic capabilities of panoramic, periapical, and Cone Beam Computed Tomography (CBCT) imaging modalities. While each modality offers distinct advantages and limitations, integrating AI with radiographic technologies introduces a new paradigm for diagnostic precision, treatment planning, and clinical outcomes in endodontics. Hence, understanding this potential requires ongoing collaboration among dental practitioners, oral and maxillofacial radiologists, and artificial intelligence researchers. Notably, AI systems demonstrate superior performance in evaluating periapical lesions in the mandible compared to those in the maxilla. Additionally, periapical radiographs offer greater sensitivity in detecting smaller lesions than panoramic radiographs, whereas CBCT provides the highest diagnostic accuracy, particularly for identifying periapical and odontogenic cystic lesions, including odontogenic keratocysts, dentigerous cysts, and periapical cysts. Importantly, the application of AI contributes to the early detection of periapical lesions, particularly those measuring as small as 1 millimeter, thereby reducing the likelihood of missed diagnoses and supporting more effective clinical decision-making.

Hence, artificial intelligence has the potential to revolutionize the diagnosis of periapical lesions by enhancing the diagnostic utility of panoramic, periapical, and Cone Beam Computed Tomography (CBCT) imaging modalities. Although each radiographic modality has inherent strengths and limitations, integrating AI with these imaging techniques offers a transformative approach to clinical diagnostics. Thus far, this synergy has facilitated greater diagnostic accuracy, optimized treatment planning, and ultimately improved patient outcomes in endodontics. To fully harness AI’s capabilities in dental radiology, sustained interdisciplinary collaboration among dental clinicians, radiologists, and artificial intelligence researchers is essential. Such cooperative efforts will ensure the responsible development and implementation of AI technologies, advancing both the precision and efficacy of periapical lesion management in contemporary dental practice.

REFERENCES