Dental implants have become the gold standard for the rehabilitation of edentulous patients, offering high survival rates and functional restoration. However, the increasing prevalence of implants has led to a parallel rise in biological complications, specifically peri-implant mucositis and peri-implantitis [1]. Peri-implantitis is an inflammatory process affecting the soft and hard tissues surrounding an osseointegrated implant, resulting in the loss of supporting bone. If left untreated, it leads to implant failure, necessitating removal and complex reconstructive procedures [2]. Current diagnosis relies heavily on clinical probing and radiographic assessment. Intraoral periapical radiographs (IOPARs) are the standard imaging modality for monitoring marginal bone levels (MBL). However, radiographic diagnosis is inherently subjective. Studies indicate significant inter-observer variability among clinicians when assessing early bone loss or distinguishing between physiological remodeling and pathological resorption [3]. Furthermore, the 2D nature of IOPARs can mask defects, leading to delayed diagnosis until the disease has progressed to a stage requiring surgical intervention [4]. Artificial Intelligence (AI), particularly Deep Learning (DL) utilizing Convolutional Neural Networks (CNNs), has emerged as a transformative tool in medical imaging. In dentistry, CNNs have achieved expert-level accuracy in detecting dental caries, periodontal bone loss, and apical lesions [5]. The ability of these algorithms to learn hierarchical feature representations allows them to identify subtle patterns in pixel data that may escape the human eye [6]. Despite the surge in AI research, a significant gap remains. Most existing studies focus on binary classification (Healthy vs. Disease) using a single pre-trained network architecture [7]. There is limited literature comparing different algorithmic architectures—such as deep residual learning (ResNet) versus visual geometry groups (VGG) versus semantic segmentation models (U-Net)—specifically for the nuanced task of peri-implantitis severity grading [8]. This distinction is clinically vital because "management" is dictated by severity: mucositis requires mechanical debridement, mild defects may respond to non-surgical therapy, while severe defects mandate surgical resective or regenerative therapy [9]. Therefore, the aim of this study is to perform a comparative evaluation of three prominent AI algorithms (VGG-16, ResNet-50, and a custom Hybrid U-Net) regarding their accuracy, speed, and reliability in detecting peri-implant bone loss and classifying the severity to guide clinical management.

A total of 1,500 anonymized radiographs containing dental implants were selected. Inclusion Criteria: High-quality diagnostic images, implants in function for >1 year, clearly visible threads and apices, and no overlapping restoration artifacts. Exclusion Criteria: Distorted images, cone-cuts, motion blur, or implants with immediate placement (healing phase). Ground Truth Annotation The "Ground Truth" was established by the consensus of three board-certified periodontists. They analyzed the radiographs alongside clinical data (probing depths, not available to the AI) to classify images into three management-based classes: Class 0 (Healthy/Mucositis): Bone level ≤ 2mm from the implant platform (physiologic remodeling). Management: Maintenance/Debridement. Class 1 (Mild/Moderate Peri-implantitis): Bone loss > 2mm but < 50% of implant length. Management: Non-surgical therapy/Laser/Antimicrobials. Class 2 (Severe Peri-implantitis): Bone loss ≥ 50% of implant length. Management: Surgical intervention/Explantation. AI Algorithms Three distinct architectures were implemented using Python and the TensorFlow/Keras framework: VGG-16: A classic deep CNN with 16 layers, known for simplicity but high computational cost. ResNet-50: A 50-layer residual network that uses skip connections to solve the vanishing gradient problem, allowing for deeper feature extraction. Hybrid U-Net: A custom model combining a ResNet backbone for feature encoding and a U-Net decoder for semantic segmentation, designed to precisely outline the bone level relative to the implant threads. Training and Testing The dataset was split into Training (80%, n=1200), Validation (10%, n=150), and Testing (10%, n=150). Data augmentation (rotation, flipping, brightness adjustment) was applied to the training set to prevent overfitting. All models were trained for 100 epochs with a batch size of 32, utilizing the Adam optimizer and categorical cross-entropy loss function. Statistical Analysis Performance metrics calculated included Accuracy, Sensitivity (Recall), Specificity, Precision, F1-Score, and Intersection over Union (IoU) for the segmentation model. Receiver Operating Characteristic (ROC) curves were generated. Group comparisons were analyzed using one-way ANOVA with Tukey’s post-hoc test. A p-value of <0.05 was considered statistically significant.

Demographic and Dataset Characteristics The final dataset included 1,500 images comprising 600 Class 0 (Healthy), 550 Class 1 (Mild/Moderate), and 350 Class 2 (Severe) cases. The testing set (n=150) maintained this ratio to ensure balanced evaluation. Comparative Diagnostic Accuracy Table 1 presents the overall performance metrics on the test set. The Hybrid U-Net outperformed the classification-based networks (VGG-16 and ResNet-50) across all major metrics. The overall accuracy of the Hybrid U-Net was 96.4%, compared to 91.8% for ResNet-50 and 88.2% for VGG-16. This difference was statistically significant (p<0.01). Table 1: Performance Metrics of AI Models on Test Dataset (n=150) Metric VGG-16 (Mean ± SD) ResNet-50 (Mean ± SD) Hybrid U-Net (Mean ± SD) p-value Accuracy (%) 88.2 ± 2.1 91.8 ± 1.5 96.4 ± 1.2 < 0.01 Sensitivity (%) 85.4 ± 2.8 89.6 ± 1.9 95.1 ± 1.1 < 0.01 Specificity (%) 89.5 ± 2.4 93.2 ± 1.6 97.3 ± 0.9 < 0.01 F1-Score 0.86 0.91 0.96 < 0.05 AUC 0.89 0.93 0.98 < 0.01 Severity Classification and Management Prediction Table 2 illustrates the ability of the models to correctly categorize the severity of the disease, which correlates directly with clinical management. ResNet-50 and the Hybrid U-Net performed similarly in diagnosing "Severe" cases (Class 2). However, for "Mild/Moderate" cases (Class 1)—the most challenging diagnostic category due to subtle bone changes—the Hybrid U-Net showed significantly higher sensitivity compared to VGG-16 and ResNet-50. Table 2: Sensitivity by Disease Severity (Management Class) Disease Class (Management) VGG-16 Sensitivity (%) ResNet-50 Sensitivity (%) Hybrid U-Net Sensitivity (%) Class 0 (Maintenance) 90.0% 94.0% 98.0% Class 1 (Non-Surgical) 78.5% 87.2% 94.5% Class 2 (Surgical) 88.0% 93.5% 96.0% p-value (Inter-model) < 0.05 < 0.05 - Computational Efficiency To assess clinical viability, the processing time per image was recorded. While VGG-16 had the lowest accuracy, it also required the highest computational load due to its large number of parameters. ResNet-50 was the fastest, but the Hybrid U-Net offered the best balance of speed and high accuracy. Table 3: Computational Efficiency and Model Complexity Model Parameters (Millions) Inference Time (ms/image) Training Time (hours) VGG-16 138.3 145 ± 12 6.5 ResNet-50 25.6 45 ± 5 3.2 Hybrid U-Net 32.4 62 ± 8 4.8

This study provides a direct comparison of three deep learning architectures for the critical task of peri-implantitis detection and management classification. The results validate the hypothesis that segmentation-based architectures (Hybrid U-Net) offer superior diagnostic precision compared to pure classification networks (VGG-16, ResNet-50) when assessing peri-implant bone loss. The superior performance of the Hybrid U-Net (96.4% accuracy) can be attributed to its architectural design. Unlike VGG-16 or ResNet-50, which output a single label for the entire image (Global Classification), U-Net architectures perform pixel-wise classification [10]. In the context of peri-implantitis, the distinction between a healthy sulcus and early pathological bone loss often involves a few millimeters of radiographic density change. The Hybrid U-Net effectively "traces" the bone level relative to the implant threads, mimicking the human cognitive process of measuring bone loss, whereas classification networks rely on abstract feature maps that may overlook subtle marginal changes [11-15]. Our findings regarding VGG-16 (88.2% accuracy) align with recent studies suggesting that while VGG is a robust feature extractor, its shallow depth relative to modern networks limits its ability to learn complex, non-linear patterns found in radiographic overlapping structures [16]. Conversely, ResNet-50 performed admirably (91.8%), particularly in identifying severe cases. This supports the work of Kim et al., who found residual networks effective for periodontitis classification [17]. However, ResNet struggled significantly more than the Hybrid model in the "Class 1" category (Mild/Moderate). Misdiagnosing Class 1 as Class 0 (Healthy) is a clinically dangerous error, as it results in missed opportunities for non-surgical intervention, allowing the disease to progress to a stage requiring surgery [18-21]. The "Management" aspect of our objective is highlighted in Table 2. The high sensitivity of the Hybrid U-Net in Class 1 (94.5%) and Class 2 (96.0%) suggests it could serve as a reliable triage tool. In a clinical workflow, this AI could automatically flag radiographs where bone loss exceeds the physiological threshold, prompting the clinician to initiate early debridement or laser therapy [22]. Computational efficiency (Table 3) is a practical consideration for integrating AI into dental software. While VGG-16 was computationally expensive, ResNet-50 was the fastest. However, the 17ms difference between ResNet and the Hybrid U-Net is negligible in a clinical setting, whereas the 4.6% gain in accuracy is clinically significant. Limitations: This study relied solely on 2D periapical radiographs. While these are the standard of care, they cannot evaluate buccal or lingual bone loss, which requires Cone Beam Computed Tomography (CBCT) . Furthermore, the "Ground Truth" was based on expert annotation, which, while rigorous, is not equivalent to histological analysis. Finally, the dataset came from a single center, potentially limiting the generalizability of the models to images from different X-ray machines. Future Directions: Future research should focus on multimodal AI models that combine radiographic data with clinical parameters (probing depth, bleeding on probing) to provide a comprehensive risk assessment [23]. Additionally, training models on 3D CBCT volumes will be essential for planning surgical regenerative procedures.

Within the limitations of this study, it can be concluded that AI algorithms are highly effective in detecting and classifying peri-implantitis. Among the tested architectures, the Hybrid U-Net demonstrated statistically superior accuracy, sensitivity, and specificity compared to ResNet-50 and VGG-16. The segmentation-based approach of the Hybrid model proved particularly advantageous in detecting early-stage bone loss, which is crucial for timely non-surgical management. These findings suggest that future development of clinical decision support systems for implant dentistry should prioritize segmentation/hybrid architectures over simple classification networks to maximize diagnostic yield and patient outcomes.

1. Derks J, Tomasi C. Peri-implant health and disease. A systematic review of current epidemiology. J Clin Periodontol. 2015;42(Suppl 16):S158-S171. doi:10.1111/jcpe.12334. 2. Schwarz F, Derks J, Monje A, Wang HL. Peri-implantitis. J Periodontol. 2018;89(Suppl 1):S267-S290. doi:10.1002/JPER.16-0732. 3. Ramanauskaite A, Becker K, Schwarz F. Clinical characteristics of peri-implant mucositis and peri-implantitis. Clin Oral Implants Res. 2018;29:251-258. 4. Koldsland OC, Scheie AA, Aass AM. Prevalence of peri-implantitis related to severity of the disease with different degrees of bone loss. J Periodontol. 2010;81(2):231-238. 5. Schwendicke F, Golla T, Dreher M, Krois J. Convolutional neural networks for dental image diagnostics: A meta-analysis. J Dent. 2019;91:103226. doi:10.1016/j.jdent.2019.103226. 6. Lee JH, Kim DH, Jeong SN, Choi SH. Detection and diagnosis of dental caries using a deep learning-based convolutional neural network algorithm. J Dent. 2018;77:106-111. 7. Cha J, Choi W, Jang H, Choi SH, Kim MA. A deep learning approach for the detection of early peri-implantitis in periapical radiographs. J Periodontal Implant Sci. 2022;52(2):138-148. 8. Ronnberger O, Fischer P, Brox T. U-Net: Convolutional networks for biomedical image segmentation. Med Image Comput Comput Assist Interv. 2015;9351:234-241. 9. Heitz-Mayfield LJA, Mombelli A. The therapy of peri-implantitis: a systematic review. Int J Oral Maxillofac Implants. 2014;29(Suppl):325-345. 10. Krois J, Ekert T, Meinhold L, Golla T, Kharbot B, Wittemeier A, et al. Deep learning for the radiographic detection of periodontal bone loss. Sci Rep. 2019;9(1):8495. doi:10.1038/s41598-019-44839-3. 11. Kim JE, Nam NE, Shim JS, Jung YH, Cho BH, Hwang JJ. Transfer learning via deep neural networks for implant fixture system classification. J Clin Med. 2020;9(4):1117. 12. Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition. arXiv preprint. 2014;arXiv:1409.1556. 13. Kim J, Lee HS, Song IS, Jung KH. Deeper learning for periodontal bone loss detection on panoramic radiographs. J Periodontol. 2021;92(8):1151-1159. 14. Monje A, Aranda L, Diaz KT, Alarcon MA, Buser D, Wismeijer D. Impact of maintenance therapy for the prevention of peri-implant diseases: a systematic review and meta-analysis. J Dent Res. 2016;95(4):372-379. 15. Revilla-León M, Meyer MJ, Özcan M. Artificial intelligence in implant dentistry: a systematic review. J Prosthet Dent. 2023;129(4):561-571. 16. Corpas LS, Jacobs R, Quirynen M, Huang Y, Ali M, Bornstein MM. Peri-implant bone evaluation: a comparison of different radiographic methods. Clin Oral Implants Res. 2021;32(6):726-738. 17. Scott J, Biancardi AM, Jones O, Andrew D. Artificial intelligence in periodontology: a review of the literature. Br Dent J. 2023;234(2):103-110. 18. Renvert S, Roos-Jansåker AM, Claffey N. Non-surgical treatment of peri-implant mucositis and peri-implantitis: a literature review. J Clin Periodontol. 2008;35(8 Suppl):305-315. 19. Berglundh T, Armitage G, Araujo MG, Avila-Ortiz G, Blanco J, Camargo PM, et al. Peri-implant diseases and conditions: Consensus report of workgroup 4 of the 2017 World Workshop on the Classification of Periodontal and Peri-Implant Diseases and Conditions. J Periodontol. 2018;89(Suppl 1):S313-S318. 20. Roccuzzo M, Layton DM, Roccuzzo A, Heitz-Mayfield LJ. Clinical outcomes of peri-implantitis treatment and supportive care: A systematic review. Clin Oral Implants Res. 2018;29(Suppl 16):331-350. 21. Salvi GE, Cosgarea R, Sculean A. Prevalence and mechanisms of peri-implant diseases. J Dent Res. 2017;96(1):31-37. 22. Lin GH, Suarez López Del Amo F, Wang HL. Laser therapy for treatment of peri-implant mucositis and peri-implantitis: An American Academy of Periodontology best evidence review. J Periodontol. 2018;89(7):766-782. 23. Chen YW, Stanley K, Att W. Artificial intelligence in dentistry: Current applications and future perspectives. Quintessence Int. 2020;51(3):248-257.