Evaluation of a no-code AI model for detecting periapical radiolucencies: impact of anatomical region on diagnostic performance Authors: Manal Hamdan, Kevin W. Yu, Marguerite Miller, Stephanie J. Sidow, Zaid Badr, Sergio E. Uribe Published: 2025 ABSTRACT BackgroundTo develop a no-code artificial intelligence (AI) model for the detection of apical radiolucent lesions and to assess how lesion location influences the model’s diagnostic performance.Methods312 periapical radiographs were retrospectively collected, each accompanied by a corresponding cone-beam computed tomography (CBCT) scan obtained within six months and a radiology report authored by board-certified oral and maxillofacial radiologists. These reports served as the reference standard, and all findings were cross-verified by the primary investigator through CBCT review. The dataset included 181 images with at least one apical radiolucent lesion and 131 lesion-free controls. Using the no-code AI platform LandingLens (Landing AI LLC, Palo Alto, CA), a diagnostic model was developed. De-identified radiographs were manually annotated with bounding boxes around lesions and automatically divided into training (70%), validation (20%), and testing (10%) subsets. Model performance was evaluated at the tooth level using sensitivity, specificity, accuracy, precision, and F1 score. To investigate the influence of anatomical sites, data were stratified by region and analyzed using chi-square and pairwise chi-square tests (α = 0.05).ResultsThe AI model demonstrated robust diagnostic performance, achieving 88.7% sensitivity, 93.0% specificity, 76.6% precision, 92.1% accuracy, and an F1 score of 0.822. Sensitivity was significantly higher in the maxilla (93.1%) compared to the mandible (83.8%; p = 0.049), whereas specificity was greater in the mandible (95.4%) than in the maxilla (90.8%; p = 0.01). No differences in performance were found between anterior and posterior teeth (p > 0.05). Regionally, the highest sensitivity was observed in the anterior maxilla (97.7%), while specificity peaked in the anterior mandible (97.9%; p < 0.05). Other diagnostic metrics did not vary significantly across regions (p > 0.05).ConclusionsThis no-code AI model demonstrated high accuracy in detecting apical radiolucent lesions and revealed that diagnostic performance can vary based on anatomical location. The model's increased sensitivity in the maxilla and higher specificity in the mandible highlight the relevance of regional anatomy in AI-assisted diagnostics. These findings support the potential of no-code AI platforms as accessible tools for building clinically meaningful diagnostic models and underscore the importance of anatomical considerations in their implementation.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12903-025-07529-3. Background Artificial intelligence (AI) refers to computer systems capable of performing tasks traditionally requiring human intelligence, such as problem-solving, decision-making, pattern recognition, and language understanding. Machine learning (ML) and deep learning (DL) are prominent subsets within AI. ML relies on algorithms that require human expertise to manually define key features for data analysis and predictions, while DL utilizes neural networks to automatically discover patterns in raw data, enabling more complex representations without manual feature extraction [1]. In clinical dentistry, AI has shown significant potential; particularly in the detection of dental caries [2], periodontal bone loss [3], and other pathologies. Its utility extends to root canal system segmentation and the identification of periapical radiolucencies, offering improved diagnostic accuracy, reduced human error, and enhanced workflow efficiency [4]. Within endodontics, the accurate detection of periapical lesions is especially critical. These lesions, often resulting from chronic inflammation due to bacterial infection and pulp necrosis, can substantially impact treatment planning and outcomes. Yet, their radiographic interpretation remains challenging, influenced by anatomical variability, imaging limitations, and observer subjectivity [5–8]. AI models developed for the detection of periapical lesions have demonstrated strong diagnostic capabilities, with pooled sensitivity and specificity reported between 0.925–0.94 and 0.852–0.96, respectively [9, 10]. Despite these promising results, many studies fail to account for variations in lesion characteristics or anatomical location—factors that may significantly affect model performance. Incorporating such stratification could help identify underperforming areas, ultimately guiding targeted refinements and enhancing the clinical relevance of AI-driven diagnostic tools [9]. Despite its promise, the widespread adoption of AI in dentistry is hindered by several challenges, including the need for programming expertise, limited interdisciplinary collaboration, and financial constraints [11]. No-code AI platforms present a compelling solution by enabling users to develop, train, and deploy AI models without the need for coding skills. These platforms are gaining traction in medicine [12], yet their application in dentistry remains relatively limited. Existing dental studies have primarily focused on tasks such as caries detection in bitewing and digital images, restoration segmentation, mesiodens detection, and classification of cervical vertebral maturation stages [13–17]. By lowering technical entry barriers, no-code platforms have the potential to democratize AI development, empowering clinicians and researchers to create diagnostic tools tailored to specific clinical needs [12, 15, 18]. This study aimed to investigate the diagnostic performance of a no-code AI model in detecting apical radiolucent lesions on periapical radiographs. Additionally, it explored whether diagnostic accuracy varied by anatomical region, including comparisons between the maxilla and mandible, as well as anterior and posterior areas. Methods Study design This retrospective study aimed to develop and evaluate the diagnostic performance of a no-code artificial intelligence (AI) model for detecting periapical radiolucent lesions on periapical radiographs. The study was determined to be non-human research by the institutional review board (IRB) of Marquette University (IRB #4474). Approval for the use of the original de-identified radiographic dataset was previously granted by the University of North Carolina (IRB #19–1430). The study followed the principles outlined in the Standards for Reporting of Diagnostic Accuracy Studies (STARD) 2015 guidelines [19]. Data collection Radiology reports from a pre-existing database of 367 cone-beam computed tomography (CBCT) volumes, acquired for endodontic evaluation between 2014 and 2019, were retrospectively reviewed. These reports were authored by one of three board-certified oral and maxillofacial radiologists. The search was conducted using specific keywords, including "apical rarefying osteitis," "apical radiolucent lesions," and "apical widening of the periodontal ligament space." Periapical radiographs from adult patients (aged 18 years and older) were included if they had been captured within a maximum of six months of the corresponding CBCT scan, to minimize the likelihood of intervening pathological changes. To ensure diagnostic reliability in the absence of histological confirmation, lesions smaller than 2 mm on the CBCT were excluded since they are not visible on periapical radiographs. CBCT scans with extensive beam hardening artifacts that compromised the visibility of the regions of interest were also excluded. Reference standard The presence or absence of periapical radiolucent lesions was established using the CBCT volumes and corresponding radiology reports as the reference standard. All findings were cross verified by the primary investigator through direct CBCT review to confirm the lesion’s presence. A final dataset of 312 periapical radiographs was curated, comprising 181 images with at least one confirmed apical radiolucency and 131 lesion-free images used as controls. Model development The no-code AI platform LandingLens (Landing AI, Palo Alto, CA, USA) was selected for this study due to its built-in object detection capabilities, which are essential for localizing periapical radiolucent lesions. Alternative platforms, such as Google’s Teachable Machine, offer only classification functionality and were therefore not suitable for the lesion-level annotation required in this project. Training and annotation All periapical radiographs, both with and without confirmed periapical radiolucencies, were uploaded to a newly created object detection project within the LandingLens platform. Lesions were manually annotated by outlining bounding boxes around each radiolucency, ensuring a 2 mm margin between the lesion periphery and the bounding box edge to standardize annotation. The dataset was automatically split into training (70%), development (20%), and testing (10%) subsets using the platform’s “Auto Split” feature, minimizing selection bias. To improve model generalization and simulate real-world variability, data augmentation was applied to the training set via the platform’s automated “Random Augment” function. Augmentation techniques included horizontal flipping, brightness and contrast adjustments and rotation. The model was trained over 40 epochs, and the confidence threshold was tuned to maximize the F1 score, ensuring a balanced trade-off between sensitivity and precision in lesion detection. Evaluation Two null hypotheses were tested, each assuming no diagnostic capability beyond chance, with any observed differences attributed to random variation:The no-code AI model cannot distinguish periapical radiolucent lesions from other radiographic structures.The model's diagnostic performance does not differ based on anatomical location. The no-code AI model cannot distinguish periapical radiolucent lesions from other radiographic structures. The no-code AI model cannot distinguish periapical radiolucent lesions from other radiographic structures. The model's diagnostic performance does not differ based on anatomical location. The model's diagnostic performance does not differ based on anatomical location. Statistical analysis Model predictions were evaluated against the reference standard and classified as true positives (TP), false positives (FP), false negatives (FN), or true negatives (TN) as shown in Fig. 1. Performance metrics including sensitivity, specificity, precision, accuracy, and F1 score were calculated at the tooth level.Fig. 1No-code AI model predictions and examples of errors. A Reference standard annotations shown before model’s predictions. B Manually annotated PARLs and model’s accurate predictions (true positive). C False positive error shown at sites #3 and #5 due to overlapping anatomical structures (maxillary sinus). D False positive error at site #30 due to enlarged bone marrow spaces. E False negative error displayed by the model’s inability to predict the lesion at site #6 due to proximity to the sinus and the presence of diffuse lesion margins No-code AI model predictions and examples of errors. A Reference standard annotations shown before model’s predictions. B Manually annotated PARLs and model’s accurate predictions (true positive). C False positive error shown at sites #3 and #5 due to overlapping anatomical structures (maxillary sinus). D False positive error at site #30 due to enlarged bone marrow spaces. E False negative error displayed by the model’s inability to predict the lesion at site #6 due to proximity to the sinus and the presence of diffuse lesion margins To address the secondary objective, the model’s diagnostic performance was evaluated across broad anatomical categories: anterior vs. posterior and maxilla vs. mandible, as well as across the four specific regions: anterior maxilla, anterior mandible, posterior maxilla, and posterior mandible. Statistical differences between anatomical locations were assessed using pairwise comparisons: Chi-square tests were applied when all cell counts were ≥ 5, Fisher’s exact test was used when any cell count was < 5, and a bootstrap-based test was employed for F1 scores. Bonferroni correction was applied to adjust for multiple comparisons to decrease the risk of type I error. The statistical significance was set at α = 0.05. Results Figure 2 summarizes the flow of the study. A total of 312 periapical radiographs encompassing 1,082 teeth were included in the analysis. Of these, 221 teeth demonstrated apical radiolucent lesions, which were manually annotated using bounding boxes. All diagnostic assessments were conducted at the tooth level. The overall performance of the no-code AI model is summarized in Table 1 through a confusion matrix, followed by the key diagnostic metrics and their formulas in Table 2.Fig. 2STARD flowchart summarizing the flow of the studyTable 1Confusion matrix summarizing overall model performanceTPFPFNTN1966025801TP True positive, FP False positive, FN False negative, TN True negativeTable 2Overall diagnostic accuracy measures of the no-code AI modelMetricFormulaValue95% CISensitivityTP/(TP + FN)88.7%(0.838, 0.922)SpecificityTN/(FP + TN)93.0%(0.911, 0.945)PrecisionTP/(TP + FP)76.6%(0.710, 0.813)Accuracy(TP + TN)/(P + N)92.1%(0.904, 0.936)F1 Score2TP/(2TP + FP + FN)0.822(0.783, 0.858) STARD flowchart summarizing the flow of the study Table 1Confusion matrix summarizing overall model performanceTPFPFNTN1966025801TP True positive, FP False positive, FN False negative, TN True negative Confusion matrix summarizing overall model performance TP True positive, FP False positive, FN False negative, TN True negative Table 2Overall diagnostic accuracy measures of the no-code AI modelMetricFormulaValue95% CISensitivityTP/(TP + FN)88.7%(0.838, 0.922)SpecificityTN/(FP + TN)93.0%(0.911, 0.945)PrecisionTP/(TP + FP)76.6%(0.710, 0.813)Accuracy(TP + TN)/(P + N)92.1%(0.904, 0.936)F1 Score2TP/(2TP + FP + FN)0.822(0.783, 0.858) Overall diagnostic accuracy measures of the no-code AI model The model’s confidence threshold was fine-tuned using the F1 score, a metric that balances both precision and recall. The optimal performance was observed at a threshold of 0.31, maximizing the F1 score. To investigate how performance varied by anatomical site, confusion matrices were generated by both broad location (anterior vs. posterior; maxilla vs. mandible) and specific region (e.g., anterior maxilla). These results are shown in Table 3.Table 3Confusion matrix by broad location and specific regionLocation/RegionTPFPFNTNAnterior62257384Posterior1343518417Mandible881917398Maxilla108418403Anterior Mandible1946188Anterior Maxilla43211196Posterior Mandible691511210Posterior Maxilla65207207 Table 3Confusion matrix by broad location and specific regionLocation/RegionTPFPFNTNAnterior62257384Posterior1343518417Mandible881917398Maxilla108418403Anterior Mandible1946188Anterior Maxilla43211196Posterior Mandible691511210Posterior Maxilla65207207 Confusion matrix by broad location and specific region Performance by broad anatomical category When comparing broad locations, sensitivity and specificity for anterior teeth were 89.9% and 93.9%, respectively, and for posterior teeth, they were 88.2% and 92.3%. As shown in Table 4, these differences were not statistically significant (p > 0.05). Similarly, the F1-scores were 0.795 for anterior teeth and 0.835 for posterior teeth (p = 0.302), indicating stable overall performance across maxillary and mandibular teeth.Table 4Diagnostic performance comparison: Anterior vs. PosteriorMetricAnterior95% CIPosterior95% CIP-valueSensitivity89.9%(0.83, 0.97)88.2%(0.83, 0.93)0.889Specificity93.9%(0.92, 0.96)92.3%(0.90, 0.95)0.421Precision71.3%(0.62, 0.81)79.3%(0.73, 0.85)0.201Accuracy93.3%(0.91, 0.96)91.2%(0.89, 0.94)0.250F1 Score0.795(0.66, 0.94)0.835(0.75, 0.93)0.302 Table 4Diagnostic performance comparison: Anterior vs. PosteriorMetricAnterior95% CIPosterior95% CIP-valueSensitivity89.9%(0.83, 0.97)88.2%(0.83, 0.93)0.889Specificity93.9%(0.92, 0.96)92.3%(0.90, 0.95)0.421Precision71.3%(0.62, 0.81)79.3%(0.73, 0.85)0.201Accuracy93.3%(0.91, 0.96)91.2%(0.89, 0.94)0.250F1 Score0.795(0.66, 0.94)0.835(0.75, 0.93)0.302 Diagnostic performance comparison: Anterior vs. Posterior In contrast, a comparison between the maxilla and mandible revealed significant differences. Sensitivity was higher in the maxilla (93.1%) compared to the mandible (83.8%; p = 0.049), whereas specificity was greater in the mandible (95.4%) than the maxilla (90.8%; p = 0.010). Other metrics were comparable (p > 0.05), as detailed in Table 5. The corresponding F1-scores were 0.815 for the maxilla and 0.830 for the mandible (p = 0.688), showing no significant differences and reflecting balanced diagnostic performance between jaws despite variability in individual metrics. Table 5Diagnostic performance comparison: Maxilla vs. MandibleMetricMandible95% CIMaxilla95% CIP-ValueSensitivity83.8%(0.77, 0.91)93.1%(0.89, 0.98)0.049*Specificity95.4%(0.93, 0.97)90.8%(0.88, 0.94)0.010*Precision82.2%(0.75, 0.90)72.5%(0.65, 0.80)0.095Accuracy93.1%(0.91, 0.95)91.2%(0.89, 0.94)0.308F1 Score0.830(0.72, 0.95)0.815(0.72, 0.91)0.688*P < 0.05 indicates statistical significance. Sensitivity and specificity were statistically significant between the maxilla and mandible, while precision, accuracy, and F1 score did not reach statistical significance Diagnostic performance comparison: Maxilla vs. Mandible *P < 0.05 indicates statistical significance. Sensitivity and specificity were statistically significant between the maxilla and mandible, while precision, accuracy, and F1 score did not reach statistical significance Performance by specific anatomical region Further stratification revealed significant variation in sensitivity and specificity across specific regions (Fig. 3 and Table 6). The highest sensitivity was observed in the anterior maxilla (97.7%), followed by the posterior maxilla (90.3%), posterior mandible (86.3%), and anterior mandible (76.0%). Specificity peaked in the anterior mandible (97.9%), followed by the posterior mandible (93.3%), posterior maxilla (91.2%), and anterior maxilla (90.3%). The F1-scores demonstrated consistent performance across sites, ranging from 0.792 in the anterior mandible to 0.841 in the posterior mandible. The anterior maxilla and posterior maxilla exhibited F1-scores of 0.796 and 0.828, respectively. No statistically significant differences were observed among all regions (p = 0.753).Fig. 3Diagnostic performance metrics heatmap across specific anatomical regionsEach cell shows the mean metric value with 95% confidence intervals in parentheses. Color intensity represents the relative magnitude of the score, with darker shades indicating higher performance. *P < 0.05 indicates statistical significance. Sensitivity and specificity were statistically significant across the four regions, while precision, accuracy, and F1 score did not reach statistical significance Diagnostic performance metrics heatmap across specific anatomical regions Each cell shows the mean metric value with 95% confidence intervals in parentheses. Color intensity represents the relative magnitude of the score, with darker shades indicating higher performance. *P < 0.05 indicates statistical significance. Sensitivity and specificity were statistically significant across the four regions, while precision, accuracy, and F1 score did not reach statistical significance Table 6Pairwise statistical comparisons of diagnostic performance metrics across specific anatomical regions, with Bonferroni-adjusted p-valuesMetricRegion ComparisonTestP-ValueP-Value (Bonferroni) Sensitivity Ant Mand vs Ant MaxFisher0.0077 0.046* Ant Mand vs Post MandFisher0.22841Ant Mand vs Post MaxFisher0.0910.5462Ant Max vs Post MandFisher0.05450.3272Ant Max vs Post MaxFisher0.25581Post Mand vs Post MaxChi-square0.44291 Specificity Ant Mand vs Ant MaxChi-square0.0014 0.0083* Ant Mand vs Post MandChi-square0.02530.1517Ant Mand vs Post MaxChi-square0.0032 0.0189* Ant Max vs Post MandChi-square0.24731Ant Max vs Post MaxChi-square0.75251Post Mand vs Post MaxChi-square0.39381 Precision Ant Mand vs Ant MaxChi-square0.1610.966Ant Mand vs Post MandFisher11Ant Mand vs Post MaxChi-square0.52991Ant Max vs Post MandChi-square0.03570.2139Ant Max vs Post MaxChi-square0.20911Post Mand vs Post MaxChi-square0.36291 Accuracy Ant Mand vs Ant MaxChi-square0.09610.5766Ant Mand vs Post MandChi-square0.08180.4909Ant Mand vs Post MaxChi-square0.05460.3277Ant Max vs Post MandChi-square0.96761Ant Max vs Post MaxChi-square0.80181Post Mand vs Post MaxChi-square0.82621 F1 Score Ant Mand vs Ant MaxBootstrap (multi)0.89841Ant Mand vs Post MandBootstrap (multi)0.1440.864Ant Mand vs Post MaxBootstrap (multi)0.31521Ant Max vs Post MandBootstrap (multi)0.07520.4512Ant Max vs Post MaxBootstrap (multi)0.22641Post Mand vs Post MaxBootstrap (multi)0.56681Abbreviations: Ant Anterior, Post Posterior, Mand Mandible, Max Maxilla* indicates statistical significance after Bonferroni correction (p < 0.05) Table 6Pairwise statistical comparisons of diagnostic performance metrics across specific anatomical regions, with Bonferroni-adjusted p-valuesMetricRegion ComparisonTestP-ValueP-Value (Bonferroni) Sensitivity Ant Mand vs Ant MaxFisher0.0077 0.046* Ant Mand vs Post MandFisher0.22841Ant Mand vs Post MaxFisher0.0910.5462Ant Max vs Post MandFisher0.05450.3272Ant Max vs Post MaxFisher0.25581Post Mand vs Post MaxChi-square0.44291 Specificity Ant Mand vs Ant MaxChi-square0.0014 0.0083* Ant Mand vs Post MandChi-square0.02530.1517Ant Mand vs Post MaxChi-square0.0032 0.0189* Ant Max vs Post MandChi-square0.24731Ant Max vs Post MaxChi-square0.75251Post Mand vs Post MaxChi-square0.39381 Precision Ant Mand vs Ant MaxChi-square0.1610.966Ant Mand vs Post MandFisher11Ant Mand vs Post MaxChi-square0.52991Ant Max vs Post MandChi-square0.03570.2139Ant Max vs Post MaxChi-square0.20911Post Mand vs Post MaxChi-square0.36291 Accuracy Ant Mand vs Ant MaxChi-square0.09610.5766Ant Mand vs Post MandChi-square0.08180.4909Ant Mand vs Post MaxChi-square0.05460.3277Ant Max vs Post MandChi-square0.96761Ant Max vs Post MaxChi-square0.80181Post Mand vs Post MaxChi-square0.82621 F1 Score Ant Mand vs Ant MaxBootstrap (multi)0.89841Ant Mand vs Post MandBootstrap (multi)0.1440.864Ant Mand vs Post MaxBootstrap (multi)0.31521Ant Max vs Post MandBootstrap (multi)0.07520.4512Ant Max vs Post MaxBootstrap (multi)0.22641Post Mand vs Post MaxBootstrap (multi)0.56681Abbreviations: Ant Anterior, Post Posterior, Mand Mandible, Max Maxilla* indicates statistical significance after Bonferroni correction (p < 0.05) Pairwise statistical comparisons of diagnostic performance metrics across specific anatomical regions, with Bonferroni-adjusted p-values Abbreviations: Ant Anterior, Post Posterior, Mand Mandible, Max Maxilla * indicates statistical significance after Bonferroni correction (p < 0.05) Bonferroni Corrected Fisher’s exact and pairwise chi-square comparisons identified significant differences in: Sensitivity between anterior maxilla and anterior mandible (p = 0.046)Specificity between anterior mandible and anterior maxilla (p = 0.008), and anterior mandible and posterior maxilla (p = 0.019). Sensitivity between anterior maxilla and anterior mandible (p = 0.046) Sensitivity between anterior maxilla and anterior mandible (p = 0.046) Specificity between anterior mandible and anterior maxilla (p = 0.008), and anterior mandible and posterior maxilla (p = 0.019). Specificity between anterior mandible and anterior maxilla (p = 0.008), and anterior mandible and posterior maxilla (p = 0.019). No statistically significant differences were found in other performance metrics (p > 0.05). Discussion Main results Periapical radiolucencies at the root apex typically indicate apical periodontitis, an inflammatory condition primarily resulting from microbial invasion of the root canal system [20]. Often asymptomatic, apical periodontitis can be challenging to detect without radiographic imaging [21]. Accurate diagnosis is crucial, as the presence of these lesions is associated with a less favorable prognosis for both nonsurgical root canal treatment and retreatment [22–24]. Artificial intelligence holds a significant promise in enhancing the detection of periapical radiolucencies, potentially improving both diagnostic accuracy and efficiency [25]. In this study, the no-code AI model achieved a sensitivity of 88.7%, demonstrating strong capability in identifying periapical lesions. Its specificity was also high at 93.0%, reflecting effective discrimination between diseased and healthy cases. However, the model's precision was only 76.6% due to a high number of false positives among the control samples. In clinical applications, false positives can erode practitioner confidence and hinder adoption. A precision of 76.6% translates to approximately one false positive for every four AI-flagged detections. While this rate is acceptable for a decision-support tool, it highlights the need for careful implementation. Within no-code frameworks, several strategies can mitigate this issue: standardizing image preprocessing and dataset curation, maintaining consistent annotation practices, including challenging negative samples during training, and adjusting confidence thresholds to reduce false detections while preserving sensitivity. Most importantly, integrating AI into a clinician-AI hybrid workflow is the optimal approach. Instead of making autonomous diagnoses, the AI model should highlight suspicious regions for expert review. This allows clinicians to benefit from high sensitivity while maintaining interpretive control and avoiding overdiagnosis. Comparison with other studies Previous studies have explored AI applications for periapical lesion detection with varying success. A study reported a sensitivity of 0.92 and precision of 0.85 using a deep learning model for lesion segmentation [26], results that closely align with the performance of the no-code model in this study. Similarly, a systematic review found pooled sensitivity and specificity values of 0.94 and 0.96, respectively, for AI-based periapical lesion detection [9], reinforcing the potential of AI tools in this diagnostic area. Model performance is known to be influenced by techniques such as data augmentation. A study used cropping and adaptive thresholding to improve image quality, achieving a model accuracy of 96.21% [27]. In the present study, the no-code model was trained with the aid of Random Augment, which introduced transformations such as horizontal flipping, rotations, and contrast adjustment to optimize performance. Another study also evaluated a deep learning model for periapical periodontitis detection, reporting an F1 score of 0.828 which is comparable to this study, though with a higher misclassification rate [28]. These comparisons suggest that no-code AI models can be just as useful as traditionally coded models. Selection of CBCT as a reference standard In this study, cone beam computed tomography (CBCT) was used as the reference standard for confirming the presence of periapical radiolucencies. CBCT provides high-resolution, three-dimensional (3D) imaging, eliminating the anatomical overlap and image distortion commonly encountered with traditional two-dimensional (2D) radiographs [21, 29]. Its diagnostic superiority for periapical lesion detection has been consistently demonstrated in the literature. Therefore, using CBCT as the reference standard strengthened the validity of our findings [30]. Variation in anatomical location Regarding anatomical location, the model demonstrated significantly higher sensitivity in the maxilla than the mandible, suggesting that dentoalveolar anatomy in the maxillary region may facilitate more accurate lesion detection. The specificity of the model was higher in the mandible, which can be attributed to the region’s fewer anatomical structures likely to mimic or obscure periapical radiolucencies. Notably, the anterior mandible exhibited significantly higher specificity than other regions (p < 0.05), likely due to the absence of confounding anatomical structures such as the maxillary sinus, mental foramen, and nasal cavity. However, when evaluated using the F1-score, no statistically significant differences were found among regions, with values ranging from 0.792 to 0.841. This suggests that despite regional variability in sensitivity and specificity, the model maintained a consistent overall diagnostic balance across diverse anatomical sites. Direct comparisons against existing studies stratifying periapical lesion detection are limited due to lack of consistency between methodological design, radiographic imaging modality, data collection, and the differences in the AI models used. When analyzing the stratified findings of the no-code AI model’s performance in periapical lesions detection, no statistically significant differences were observed in any performance metric between the anterior and posterior teeth. Similarly, in a study using You Only Look Once version 4 (YOLOv4) and Faster regions with convolutional neural network (R-CNN) no significant differences were reported in performance metrics for detecting periapical lesions on periapical radiographs when comparing anterior teeth, premolars, and molars [31]. In contrast, research using a custom convolutional neural network (CNN) for panoramic radiographs found higher sensitivity but lower specificity in molars [4]. Additionally, when a periapical lesion detection algorithm was employed for CBCT scans, sensitivity was reported to be lower in lateral incisors, canines, and third molars than in other tooth groups [32]. These findings suggest that, despite differences in methodologies, tooth type may not have a significant impact on periapical lesion detection in periapical radiographs but may have limited effects in other imaging modalities. To control for or further analyze the effects of these anatomical confounders, some studies have limited their analysis to either the maxilla or mandible [33, 34]. For example, when detecting maxillary periapical lesions greater than 2 mm on CBCT images, a Siamese Concatenated Network alongside DenseNet-121 or VGG-16 achieved a sensitivity of 0.5910 (VGG-16) and specificity of 0.9239 (DenseNet-121) [33]. By comparison, the no-code AI model exhibited higher sensitivity (0.931) in the maxilla, but similar specificity (0.908). In a periapical radiograph study focused on mandibular periapical lesion detection based on periapical index (PAI) scores of 1 and 2 versus 3, 4, and 5, YOLOv3 achieved a sensitivity of 92.1%, similar to the no-code AI model, but had lower specificity 76.0% than the no-code model in this study. These differences may result from the subjective nature of PAI scoring, quantification by PAI score on a per-root basis by YOLOv3 versus detection of a lesion on a per-tooth basis by the no-code model [34]. Region specific comparisons revealed that the no-code model had higher sensitivity in the anterior maxilla but lower sensitivity in the anterior mandible compared to YOLOv4 and Faster R-CNN, while specificity in the anterior mandible was higher with the no-code model. In the posterior regions, the no-code model’s performance was comparable to YOLOv4 but lower than Faster R-CNN, which generally demonstrated higher performance metrics in all areas [31]. When compared to another study by the same research group utilizing only Faster R-CNN, the aforementioned region-specific findings were similar, however, the no-code AI model’s performance in the posterior maxilla and posterior mandible appeared to exhibit lower specificity and accuracy compared to the Faster R-CNN model [34]. Despite the lack of significant differences in periapical lesion detection between anatomical regions using YOLOv4 and Faster R-CNN, the no-code AI model demonstrated notable variations in performance based on these regions [31, 35]. AI models offer promising support, particularly for less experienced clinicians [15, 36]. Studies have shown that AI can enhance sensitivity in lesion detection and reduce diagnostic time for novice practitioners. However, certain anatomical regions -such as the maxillary anterior- may still require additional oversight to reduce false positives and ensure diagnostic accuracy [4, 15, 25, 37]. Overall, the integration of AI into clinical workflows has the potential to improve diagnostic consistency and accuracy, particularly in settings with limited expertise or high time pressures [36]. Limitations While our no-code AI model demonstrated promising performance, several limitations must be considered. First, the relatively small sample size and reliance on data from a single institution limit the generalizability of the findings. Multi-center studies with larger and more diverse datasets are needed to better assess the model’s robustness across different clinical settings. Although the dataset used in this study was of high quality, it may not reflect the variability typically encountered in real-world scenarios, which could affect the model's scalability. The absence of an external testing also limits the assessment of the model’s performance in broader, other scenarios. Another important limitation is the use of CBCT and radiology reports as the sole reference standard. While CBCT is widely regarded as the gold standard for detecting periapical radiolucencies, integrating clinical findings or histological confirmation would strengthen the model’s diagnostic reliability. However, these factors were beyond the scope of the present study. There are also inherent limitations associated with no-code AI platforms themselves. These platforms often provide limited control over training parameters, making it difficult to determine optimal stopping points and increasing the risk of underfitting or overfitting. Moreover, most no-code tools also lack support for DICOM files and are generally restricted to 2D image classification, limiting their applicability in clinical environments where 3D imaging and more complex tasks like segmentation or object detection are required. The black-box nature of these platforms reduces transparency, as users cannot inspect intermediate layers or decision rules. However, human diagnostic decisions also lack complete transparency, as clinicians cannot always articulate their reasoning processes [38].This lack of interpretability can affect clinical trust and complicate regulatory approval, as agencies such as the United States FDA increasingly require evidence of model explainability. Several strategies can enhance interpretability within no-code frameworks: visualization tools such as heatmaps or attention maps that highlight contributing regions, iterative review of errors to guide dataset refinement, and thorough documentation of data sources and preprocessing steps. Ultimately, what matters most is establishing clear responsibility for diagnostic decisions, ensuring that clinicians retain final interpretive authority. These approaches can partially mitigate the black-box limitation, supporting more trustworthy clinical use of no-code AI. Lastly, subscription-based pricing structures may pose financial barriers for individual clinicians, smaller practices, or academic institutions. Conclusion This study indicates that the developed no-code AI model can effectively detect periapical radiolucencies on periapical radiographs, demonstrating high sensitivity and specificity. The accessibility of no-code platforms offers the potential to empower clinicians to develop and implement diagnostic tools without the need for extensive programming expertise. However, further clinical research is essential. Validation using larger, more diverse datasets is needed to confirm these findings and to assess the broader clinical applicability of no-code AI in routine dental practice. Supplementary Information Supplementary Material 1. Supplementary Material 1.