OralVision: A Deep Learning Framework for Multi-Class Oral Disease Classification

Introduction

Oral diseases continue to be one of the most common non-communicable diseases of the world. In recent years, a Global Burden of Disease (GBD) analysis estimated that lip and oral cavity cancer represents a substantial and growing proportion of the total burden of disease from cancer, with differences across regions, sexes, and exposures to risk factors [1]. In addition to malignant diseases, many other common diseases occur, affecting large numbers of people and, when identified early, are known to have downstream treatment cost and disease progression implications such as dental caries, gingivitis, calculus deposition, hypodontia, oral ulceration, and tooth discoloration. Conventional diagnosis of these conditions has relied on manual, visual examination by a dental professional, which is open to inter-observer variation and limited by the geographical distribution of specialists available [2] and can result in an important diagnostic bottleneck in low-resource and remote areas.

To overcome this bottleneck, artificial intelligence, specifically convolutional neural networks (CNNs), has been suggested, achieving classification accuracy rates often surpassing 90% in many oral and dental imaging tasks and as part of the increasing interest in the use of generative AI in dentistry [3]. However, automated, image-based screening is potentially useful because it can be carried out without the presence of a specialist at the time of image capture, and before formal referral, which could be made if needed.

Solutions to this challenge have been developed sequentially: from handcrafted-feature machine-learning processes to deep convolutional architectures, and, more recently, to transfer learning from large pretrained backbones, which are discussed in detail in Section 2. This has tended to increase accuracy in the reported results and decrease the amount of feature engineering, but hasn’t addressed some common methodological drawbacks that have been observed in the literature.

The present study is motivated by four of these limitations.

• First, models are trained and tested almost always on a single dataset and a single train/test split, and cross-dataset or testing on externally sourced data is rarely reported as standard practice [4, 5].

• Second, data leakage between the training and evaluation partition may cause an overly optimistic evaluation of the model’s internal performance, when the images are not the same or when they have finite variation in their attributes [6, 7].

• Third, oral disease taxonomies are often imbalanced, that is, the conditions that occur less often (e.g., tooth discoloration) are systematically underrepresented compared with the more common conditions (e.g., tooth decay), and mechanisms designed to correct imbalances are often presumed to operate as intended but not actually tested [8].

• Fourth, interpretability tools such as Grad-CAM are increasingly used to support trust in these models, but their outputs are qualitative and are frequently presented without the caveats needed to prevent over-interpretation [9, 10].

Only limitations directly supported by the literature reviewed in Section 2 are listed here.

Overall, this literature suggests that while the accuracy on an internal test set of an oral disease classifier might be near its maximum, the real-world value is not established: the behavior of a model on data not provided by the model developer is not known. The model’s sensitivity to architectural and training decisions is unknown and the model’s interpretability is not known. We do not know of any paper reporting a classification head that is minimal and statistically justified, as well as being factorized and position-preserving for the task of oral disease classification.

We address this gap with OralVision, which combines an EfficientNet-B0 backbone, factorized Coordinate Attention, and a single-layer linear classification head. EfficientNet-B0 was selected as a comparatively compact pretrained backbone, while Coordinate Attention was included to preserve height- and width-specific information in the feature map before global average pooling. These choices provide the architectural motivation for the model and their empirical contribution is assessed through the ablation study in Section 4.5 rather than assumed in advance. The classification head is kept deliberately simple, and its design is examined separately from the classifier-family comparison in Section 4.5.

This paper has the following contributions:

• An exact-duplicate and label-conflict-corrected dataset acquired in a real-world hospital setting

• A complete ablation study on the individual role of Coordinate Attention, curriculum learning, hard-example cascading, and classifier-head depth, with a full training budget.

• A post hoc comparison of families of classifiers, based on a statistically sound approach (bootstrap confidence intervals and paired McNemar’s tests)

• A strict evaluation of the test set, using an external dataset different from the one used for training, and without any fine-tuning or parameter updates and

• A qualitative Grad-CAM interpretability analysis plus an explicit computational-complexity evaluation.

The rest of the paper describes related work (Section 2), the proposed model and methodology (Section 3), the experimental data, process, and results (Section 4), a discussion of the results and a comparison with the state of the art (Section 5), and concludes the research (Section 6).

Complete Article

The complete article, including all figures, tables, equations and algorithms, is available in the official publication PDF.

View Full PDF

Conclusion

Although there are compelling reasons for moving toward computer-based methods for analyzing photographs to diagnose oral disease, the kind of data used to train and validate the algorithms is almost always from the development set, which is limited to a single source, and the question remains whether reported accuracy is a true representation of the algorithms’ diagnostic capacity or an artifact of limited, single-source evaluation. This paper introduces a new method, named OralVision, based on combining an EfficientNet-B0 backbone with a factorized Coordinate Attention module and a single-layer linear classification head, which is trained on a leakage-audited, deduplicated dataset from Holy Family Hospital in Rawalpindi, Pakistan, and tested using a full training-budget ablation study, a statistically grounded comparison of classifiers in the family, and a strict, zero-shot evaluation on an independent, publicly documented external dataset. The primary internal test checkpoint achieved accuracy 0.995257 and macro-F1 0.984145. Table 6, Variant 6 (Naive Training), which omits Coordinate Attention, curriculum, and class weighting, reports accuracy 0.9952569 and macro-F1 0.9840564. These values are close but not identical at full precision, although they round alike at four decimal places. They represent different configurations and are reported separately. Five-run results were 99.24%±0.21% accuracy and 97.68%±0.51% macro-F1. The full-budget ablation does not provide evidence that Coordinate Attention improves performance: the no-attention variant scored higher than the Full Model baseline, as did the no-curriculum and no-hard-example-cascade variants. The classifier-family comparison found no significant difference between the separately evaluated Logistic Regression classifier and the joint head (McNemar’s p=0.617). Table 8 reports that comparison on frozen embeddings, while the primary checkpoint uses the end-to-end linear head described in Section 3.1.4. Without optimization on the independent external test set, macro-F1 was 0.79-0.82, with Tooth Discoloration recall of 0.16-0.28. The internal-versus-external gap demonstrates that high internal performance does not guarantee generalization and should be assessed directly.

Building on these findings, the most immediately valuable next steps are: extending external testing to a genuinely multi-center design beyond the single external dataset used here. A dedicated four-way (sampler × loss) decomposition experiment to test the possible class-weighting/sampler interaction suggested by the present ablation. Re-running the six remaining ablation configurations and a multi-seed ablation study at full training budget. Targeted data collection to increase Tooth Discoloration representation, quantitative interpretability evaluation against expert-annotated lesion boundaries and conformal-prediction-based data cleaning as a complementary safeguard against residual label noise.

Data Availability

The internal training dataset consists of private clinical oral photographs obtained from Holy Family Hospital and is not publicly available. The independent external test dataset is a publicly available Roboflow collection [40].

References

1. M. Hu, H. Chen, R. Wang, R. Zhang, J. Yao, X. Qiu, and J. Guo, “Global, regional, and national burden of lip and oral cavity cancer and its attributable risk factors from 1990 to 2021: An analysis of the Global Burden of Disease Study 2021, ” International Journal of Clinical Pharmacy, vol. 48, no. 1, pp. 93-106, 2026, doi: 10.1007/s11096-025-01961-9
2. H. Li et al., “Application of artificial intelligence in oral health management: Challenges and opportunities, ” Frontiers in Medicine, vol. 13, Art. no. 1700529, 2026, doi: 10.3389/fmed.2026.1700529
3. Z. Ma, C. Du, Q. Lao, and X. Xie, “Generative artificial intelligence: Applications and future prospects in dentistry, ” International Dental Journal, vol. 76, no. 1, Art. no. 108276, 2026, doi: 10.1016/j.identj.2025.108276
4. A. Musa, R. Prasad, P. Onwualu, and M. Hernandez, “A systematic review of cross-population shifts in medical imaging analysis with deep learning, ” Big Data and Cognitive Computing, vol. 10, no. 3, Art. no. 76, 2026, doi: 10.3390/bdcc10030076
5. B. Van Calster et al., “Evaluation of performance measures in predictive artificial intelligence models to support medical decisions: Overview and guidance, ” The Lancet Digital Health, vol. 7, no. 12, Art. no. 100916, 2025, doi: 10.1016/j.landig.2025.100916
6. A. Apicella, F. Isgr`o, and R. Prevete, “Don't push the button! Exploring data leakage risks in machine learning and transfer learning, ” Artificial Intelligence Review, vol. 58, no. 11, Art. no. 339, 2025, doi: 10.1007/s10462-025-11326-3
7. A. Thakur, T. Zhu, V. Abrol, J. Armstrong, Y. Wang, and D. A. Clifton, “Data encoding for healthcare data democratization and information leakage prevention, ” Nature Communications, vol. 15, Art. no. 1582, 2024, doi: 10.1038/s41467-024-45777-z
8. D. Zhou, S. Gao, and X. Huang, “Strategies for class-imbalanced learning in multi-sensor medical imaging, ” Sensors, vol. 26, no. 6, Art. no. 1998, 2026, doi: 10.3390/s26061998
9. A. Chaddad, Y. Hu, Y. Wu, B. Wen, and R. Kateb, “Generalizable and explainable deep learning for medical image computing: An overview, ” Current Opinion in Biomedical Engineering, vol. 33, Art. no. 100567, 2025, doi: 10.1016/j.cobme.2024.100567
10. A. Demircioglu, “Reproducibility and interpretability in radiomics: A critical assessment, ” Diagnostic and Interventional Radiology, vol. 31, no. 4, pp. 321-328, 2025, doi: 10.4274/dir.2024.242719
11. H. Malik, A. Naeem, D. Jeong, R. A. Naqvi, and S.-W. Lee, “QML: Quantum machine learning and vanilla GAN-based classification of dental caries using bitewing radiographs, ” Cluster Computing, vol. 29, Art. no. 355, 2026, doi: 10.1007/s10586-026-06116-7
12. M. R. Salmanpour, A. H. Pouria, S. Barichin, Y. Salehi, S. Falahati, I. Shiri, M. Oveisi, and A. Rahmim, “PySERA: Open-source standardized Python library for automated, scalable, and reproducible handcrafted and deep radiomics, ” Computer Methods and Programs in Biomedicine, vol. 284, Art. no. 109463, 2026, doi: 10.1016/j.cmpb.2026.109463
13. A. Ahmed, “Optimized ResNet-18 architecture for multi-class oral diseases classification, ” Karbala International Journal of Modern Science, vol. 12, no. 1, Art. no. 8, 2026, doi: 10.33640/2405-609X.3446
14. R. M. dos Santos, K. F. Gameiro, V. M. da Silva Ferreira, S. C. Brum, A. S. Carlos, A. M. Moreno, and S. T. A. de Mello, “Development and evaluation of a deep learning model for dental image classification, ” Research, Society and Development, vol. 15, no. 7, Art. no. e6615751427, 2026, doi: 10.33448/rsd-v15i7.51427
15. A. Z. Abdian, M. M. Javidi, N. Mansouri, F. Mehranfar, and S. Cheperli, “MultiDentNet: A unified deep learning framework for multi-class dental condition screening and preliminary oral lesion triage, ” Scientific Reports, vol. 16, Art. no. 21815, 2026, doi: 10.1038/s41598-026-52264-6
16. M. Oktiana, M. Fitria, H. Habibie, K. Saddami, Y. Elma, H. Rahayu, S. Janura, R. Putri, and R. Novita, “Performance analysis of ResNet-50, EfficientNet B4, and MobileNet V2 for dental caries classification, ” JOIV: International Journal on Informatics Visualization, vol. 10, no. 1, pp. 10-18, 2026, doi: 10.62527/joiv.10.1.3713
17. H. S. Marie, M. Elbaz, R. S. Soliman, and A. A. Elkhatib, “DentoMorph-LDMs: Diffusion models based on novel adaptive 8-connected gum tissue and deciduous teeth loss for dental image augmentation, ” Scientific Reports, vol. 15, Art. no. 27268, 2025, doi: 10.1038/s41598-025-11955-2
18. W. Wang, R. Yang, J. Xiao, Y. Wang, and Z. Chen, “Mining hierarchical oral disease prototypes: Oral disease diagnosis using a novel dynamic compositional prototype learning framework, ” Preprints, version 2, 2025, doi: 10.20944/preprints202504.0003.v2
19. D. A. Ali and H. T. Sadeeq, “An interpretable deep learning framework for multi-class dental disease classification from intraoral RGB images, ” Statistics, Optimization & Information Computing, vol. 14, no. 6, pp. 3380-3397, 2025, doi: 10.19139/soic-2310-5070-2880
20. A. B. George, S. Bathini, and N. S. R., “Deep learning for oral health: Benchmarking ViT, DeiT, BEiT, ConvNeXt, and Swin Transformer, ” arXiv preprint arXiv:2509.23100, 2025.
21. F. Y. I. Asiri, “Automated classification of enamel caries from intraoral images using deep learning models: A diagnostic study, ” Journal of Clinical Medicine, vol. 14, no. 24, Art. no. 8959, 2025, doi: 10.3390/jcm14248959
22. J. Michelin, A. Nooruddin, A. Khalid, S. M. F. Ahmed, B. King, M. M. Shaikh, A. J. T. Neto, and F. Umer, “Deep learning models for the detection of dental-findings and tooth-types using video data, ” BMC Oral Health, vol. 26, Art. no. 1502, 2026, doi: 10.1186/s12903-026-08861-y
23. L. Ren and J. Chen, “A dual-branch deep learning framework with explainability for dental caries classification using intra-oral photographs and radiographs, ” Journal of Imaging, vol. 12, no. 5, Art. no. 207, 2026, doi: 10.3390/jimaging12050207
24. K. M. Desai et al., “Screening of oral potentially malignant disorders and oral cancer using deep learning models, ” Scientific Reports, vol. 15, Art. no. 17949, 2025, doi: 10.1038/s41598-025-02802-5
25. P. Liu and K. Bagi, “A tailored deep learning approach for early detection of oral cancer using a 19-layer CNN on clinical lip and tongue images, ” Scientific Reports, vol. 15, Art. no. 23851, 2025, doi: 10.1038/s41598-025-07957-9
26. A. Mohapatra and R. Dash, “A unified framework for oral squamous cell carcinoma detection using machine learning and deep learning techniques, ” Discover Computing, vol. 29, Art. no. 169, 2026, doi: 10.1007/s10791-026-10054-8
27. N. V. S. S. Vissamsetti and G. Shivakanth, “A deep learning-based image processing framework for oral lesion classification, ” Engineering, Technology & Applied Science Research, vol. 16, no. 2, pp. 33266-33271, 2026, doi: 10.48084/etasr.15583
28. T. D. Pham and S. Al-Hebshi, “Classification of pediatric dental diseases from panoramic radiographs using natural language transformer and deep learning models, ” Frontiers in Artificial Intelligence, vol. 9, Art. no. 1754498, 2026, doi: 10.3389/frai.2026.1754498
29. B. Ayhan, E. Ayan, and S. Ats"u, “Detection of dental caries under fixed dental prostheses by analyzing digital panoramic radiographs with artificial intelligence algorithms based on deep learning methods, ” BMC Oral Health, vol. 25, Art. no. 216, 2025, doi: 10.1186/s12903-025-05577-3
30. A. Qureshi, Q. U. A. Aini, E. Solomin, N. Ahmad, Z. Uddin, and A. H. Milyani, “Dental arch classification using deep learning and federated learning on DenPAR radiographs, ” Scientific Reports, 2026, doi: 10.1038/s41598-026-62329-1
31. J. Xu, H. Gao, and Z. Wang, “KC-UNet: Enhancing U-Net with KAN and CBAM for medical image segmentation, ” IEEE Access, vol. 13, pp. 153788-153797, 2025, doi: 10.1109/ACCESS.2025.3605148
32. M. Ferdous, S. Mahmud, and M. E. Z. Shimul, “MedNet: A lightweight attention-augmented CNN for medical image classification, ” Scientific Reports, vol. 15, Art. no. 41936, 2025, doi: 10.1038/s41598-025-25857-w
33. M. Tan and Q. Le, “EfficientNet: Rethinking model scaling for convolutional neural networks, ” in Proc. 36th Int. Conf. Mach. Learn. (ICML), vol. 97, pp. 6105-6114, 2019.
34. I. Ahammad, “Generalization over accuracy: A cross-dataset, explainable, and federated learning framework for Parkinson's disease detection, ” Digital Health, vol. 12, Art. no. 20552076261467837, 2026, doi: 10.1177/20552076261467837
35. A. Musa, R. Prasad, and M. Hernandez, “Addressing cross-population domain shift in chest X-ray classification through supervised adversarial domain adaptation, ” Scientific Reports, vol. 15, Art. no. 11383, 2025, doi: 10.1038/s41598-025-95390-3
36. R. Bhuria, S. Gupta, R. M. Ghoniem, J. Singh, S. Rani, B. M. Taye, and S. Bharany, “A transfer learning-based approach for automated bone fracture classification in X-ray imaging, ” Therapeutic Advances in Musculoskeletal Disease, vol. 18, Art. no. 1759720X251405099, 2026, doi: 10.1177/1759720X251405099
37. Q. Hou, D. Zhou, and J. Feng, “Coordinate attention for efficient mobile network design, ” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2021, pp. 13713-13722, doi: 10.1109/CVPR46437.2021.01350
38. Z. Wei, F. Yang, K. Zhong, and L. Yao, “PCB-YOLO: Enhancing PCB surface defect detection with coordinate attention and multi-scale feature fusion, ” PLOS ONE, vol. 20, no. 6, Art. no. e0323684, 2025, doi: 10.1371/journal.pone.0323684
39. B. Ozdemir and I. Pacal, “An innovative deep learning framework for skin cancer detection employing ConvNeXtV2 and focal self-attention mechanisms, ” Results in Engineering, vol. 25, Art. no. 103692, 2025, doi: 10.1016/j.rineng.2024.103692
40. X. Zhan, Q. Xu, Y. Zheng, G. Lu, and O. Gevaert, “Reliability-enhanced data cleaning in biomedical machine learning using inductive conformal prediction, ” PLOS Computational Biology, vol. 21, no. 2, Art. no. e1012803, 2025, doi: 10.1371/journal.pcbi.1012803
41. P. Udomluck, W. Cholamjiak, J. Inpun, and W. Waratamrongpatai, “Multi-task deep learning model for automated detection and severity grading of lumbar spinal stenosis on MRI: Multi-center external validation, ” Diseases, vol. 14, no. 1, Art. no. 32, 2026, doi: 10.3390/diseases14010032
42. A. Alqhatani, T. K. S. R. Babu, T. R. Mahesh, S. B. Khan, O. Saidani, and M. T. Quasim, “Automated classification and explainable AI analysis of lung cancer stages using EfficientNet and gradient-weighted class activation mapping, ” Frontiers in Medicine, vol. 12, Art. no. 1625183, 2025, doi: 10.3389/fmed.2025.1625183
43. M. Akram, M. Adnan, S. F. Ali, J. Ahmad, A. Yousef, T. A. N. Alshalali, and Z. A. Shaikh, “Uncertainty-aware diabetic retinopathy detection using deep learning enhanced by Bayesian approaches, ” Scientific Reports, vol. 15, Art. no. 1342, 2025, doi: 10.1038/s41598-024-84478-x
44. M. Moharrami, E. Vahab, M. Bagherianlemraski, G. Hemmati, S. Singhal, C. Quinonez, F. Schwendicke, and M. Glogauer, “Deep learning for detecting dental plaque and gingivitis from oral photographs: A systematic review, ” Community Dentistry and Oral Epidemiology, vol. 53, no. 6, pp. 617-632, 2025, doi: 10.1111/cdoe.70001
45. P. Mirfendereski, G. Y. Li, A. T. Pearson, and A. R. Kerr, “Artificial intelligence and the diagnosis of oral cavity cancer and oral potentially malignant disorders from clinical photographs: A narrative review, ” Frontiers in Oral Health, vol. 6, Art. no. 1569567, 2025, doi: 10.3389/froh.2025.1569567
46. A. Tariq, S. I. Khan, M. Imran, and H. Dawood, “Enhanced Grad-CAM segmentation for multi-class brain tumor classification, ” in Proc. 4th Int. Conf. Communication, Computing and Digital Systems (C-CODE), 2025, doi: 10.1109/C-CODE67372.2025.11204150
47. M. H. Mujtaba, M. Imran, S. Irfan, U. Muhammad, and H. Dawood, “TamperNet: A spatio-temporal deep learning framework for video manipulation detection, ” Neural Processing Letters, 2026, doi: 10.1007/s11063-026-11885-8
48. Final Year Project, “Dental Diseases Classification Dataset, ” Roboflow Universe, Roboflow, 2025. [Online]. Available: https://universe.roboflow.com/final-year-project-gifyj/dental-diseases-classification. [Accessed: Sep. 6, 2026].