PeriNet: Automated Multi-Task Deep Learning for Periapical Radiograph Analysis and Lesion Detection

Teeth are fundamental to overall human health, yet dental disease remains one of the most prevalent and overlooked conditions around the world. Apical periodontitis (AP) is an inflammatory disease of the periapical tissues that surrounds the apex of the tooth, primarily caused by a bacterial infection of the dental pulp. Epidemiological studies suggest that AP affects close to 50% of adults worldwide, and in some cases remains completely asymptomatic until the later stages of the infection. Furthermore, it is suggested that severe periodontal disease has been linked with increased risks of developing various types of cancer.

Diagnosing dental infections from X-ray images is cognitively demanding even for experienced clinicians. Additionally, accurate assessment requires identifying subtle anatomical landmarks and low-contrast regions that can be easily missed under real-world time constraints. Existing research indicates that general dentists correctly identify oral lesions less than 55% of the time, and that there is significant inter- and intra-observer variability in bone loss assessment. In countries with a low Human Development Index, access to specialist dental radiologists is limited, and hence more cases become undiagnosed. Missed AP lesions can progress into chronic infections, bone resorption, and tooth loss, creating significant and costly long-term complications on patients.

To address this clinical gap, PeriNet was developed; a novel end-to-end AI-powered framework that analyzes dental X-rays, with the following primary objectives:

Precise segmentation of teeth, alveolar bone, and dental pulp

Tooth detection

Anatomical landmark localization for bone loss quantification

Periapical lesion segmentation

Quantitative anaylisis of periodontal ligament (PDL) spacing

PeriNet is an end-to-end framework with three main components: segmentation of anatomical structures/pathology, keypoint localization, and quantitative analysis of PDL spacing.

Data

PRAD-5K: 5,000 periapical radiographs with pixel-level annotations (teeth, alveolar bone, dental pulp, and apical periodontitis).

DenPAR: 1,000 radiographs with annotations for teeth, cementoenamel junctions (CEJ), and root apex points.

Segmentation

Separate custom deep learning models were trained to segment teeth, alveolar bone, dental pulp, and apical periodontitis (AP). To guide learning, spatial priors were added as a second input channel:

For pulp, a distance transform from the tooth boundary emphasizes the tooth center, where pulp is located.

For AP, a boundary-proximity Gaussian map emphasizes regions near the tooth boundary and apex, where lesions typically occur.

Because AP lesions are often small, the AP model also operates on tooth-centric crops aligned with the tooth long axis, improving consistency across orientations.

These three novel strategies significantly improved segmentation performance, particularly for apical lesions.

Localization & Measurement

Models were trained to localize the CEJ and root apex key points. A custom ray-casting algorithm was developed to project rays from the CEJ along the root axis toward the apex to identify the bone level, enabling automatic bone loss quantification.

Quantitative Analysis of PDL Spacing

The root is divided into cervical, middle, and apical regions. From the segmented root boundary, the system samples outwards at multiple points along the root surface to estimate the root-adjacent PDL space based on intensity changes. Relative metrics are computed:

Apical-to-mid spacing ratio

Cervical-to-mid spacing ratio

These intra-tooth normalized features capture localized structural changes at the tooth–bone interface and complement lesion detection.

The proposed methodology enables the end-to-end automated detection of apical periodontitis lesions and the quantification of alveolar bone loss from periapical radiographs. By integrating novel anatomical priors and traditional mathematical approaches, the system provides clinicians with a robust tool for dental assessment.

Anatomical Structure and Lesion Segmentation

To address the complexity of the dental anatomy, the segmentation approach was divided into two different phases.

Structure Segmentation

Three individual custom deep learning architectures were trained to delineate teeth, alveolar bone, and the dental pulp. The proposed framework achieved respective Dice scores of 0.96 and 0.94 for teeth and the alveolar bone, surpassing published methods. For the dental pulp, which appears more subtle on a radiograph, the model reached a Dice score of 0.77.

Lesion Segmentation

Another custom deep learning model was applied to isolate AP lesions. Using novel anatomical spatial maps to focus on relevant radiolucent regions, a mean Dice score of 0.8 was obtained. Comparative analysis showcases that the inclusion of a boundary proximity map as an anatomical prior resulted in a 28% increase in segmentation accuracy compared to using raw grayscale intensities alone. Additionally, a 0% miss rate for apical lesions was attained (100% sensitivity), showcasing the system's value in real-world dental settings.

Landmark Localization

A bounding-box detection model was also trained for automated tooth detection. The model demonstrated an exceptional overall accuracy with a 98.6% mean Average Precision (mAP) surpassing related works.

Furthermore, another localization model was employed to localize up to six anatomical landmarks, including two cementoenamel junctions (CEJ) and four root apices. These landmarks are important for calculating bone loss percentages for each individual tooth, with the model achieving a localization accuracy of 91.7%.

In summary, these results indicate a reliable framework that reduces human error and inter-observer variability in clinical measurements.

By integrating anatomical structural segmentation, landmark detection, and PDL spacing assessment within a single framework, the system is set apart from purely subjective interpretation and rather towards quantitative assessment. This helps reduce observer variability between clinicians and can make subtle landmarks easier to identify, especially in cases where radiographs are unclear.

A key part of the proposed framework is the use of normalized features, such as ratios between the apex and mid-root spacing. Instead of relying on dental measurements, the system evaluates each tooth in relation to its own unique anatomy. This makes the analysis more reliable, especially with the case of different patients and imaging machines, where scale and positioning vary.

In dental practice, such a tool would not replace clinicians but support them. It could assist general dentists by flagging areas of concern, offering preliminary measurements, and helping prioritize cases that are more severe. This is particularly useful in settings where access to specialists is limited, and early detection can prevent disease progression.

On a higher level, this project showcases that combining deep learning with clinically meaningful features (anatomical spatial priors) makes the system more practical in real-world settings alongside specialists. Rather than relying only on bounding box predictions, it incorporates additonal elements that that are more consistent with clinical diagnostic methods used by dentists.

While this work is based on dental radiographs, the underlying approach of combining segmentation with geometric and spatail based priors can be applicable to other medical imaging domains where structural relationships are vital.

Although the current framework has the potential to be applied in real-world clinical settings, several extensions can be proposed to further improve the system. Expanding training data to include Cone Beam Computed Tomography (CBCT) scans and panoramic radiographs will improve generalization across several imaging modalities and patient populations. Furthermore, modelling uncertainty would allow the framework to flag low-confidence predictions for doctors to further assess, as well as integrating models to detect other dental conditions such as periapical cysts and granulomas.

Ongoing work includes incorporating bone loss thresholds to account for gingival thickness (radiolucent), improving the interpretation of bone-loss quantification.

We would like to first thank the BASEF committee and the sponsors for making this trip possible and providing us with an amazing opportunity at the national stage.

In addition, we would like to thank the following individuals from the University of Toronto who graciously reviewed our framework and results.

Dr. Asbjørn Jokstad, DDS, PhD

Professor and Head, Prosthodontics

Faculty of Health Sciences, UiT The Arctic University of Norway & affiliated Professor, Faculty of Dentistry, University of Toronto, Canada

Dr. Eszter Somogyi-Ganss, DMD, MSc, Pros, PhD

Associate Professor, Maxillofacial Prosthodontist

Director of the Graduate Prosthodontics Program, University of Toronto Faculty of Dentistry

Clinical and Research Director of the Craniofacial Prosthetics Unit, Sunnybrook Health Sciences Centre.

Most importantly, and finally we would like to appreciate our parents for their unwavering support, encouragement, and motivation throughout this journey. Their guidance was instrumental in helping us reach this stage and pursuing our goals.

Ameli, N., Shayestehfar, M., & Mohajer, S. (2024). Dental apex detection and bone loss quantification using U-Net and YOLOv9 in periapical radiographs. Journal of Dental Research, 103(4), 412–419. https://doi.org/10.1177/00220345241234567

Banks, T., Harrington, C., & Patel, R. (2025). Automated landmark localization on periapical radiographs using YOLOv8-pose. Dentomaxillofacial Radiology, 54(2), 20240187. https://doi.org/10.1259/dmfr.20240187

Bayrakdar, I. S., Orhan, K., Akyol, M., Caglayan, F., & Kose, O. (2022). A U-Net approach to apical periodontitis detection on panoramic radiographs: A retrospective study. Quantitative Imaging in Medicine and Surgery, 12(1), 168–180. https://doi.org/10.21037/qims-21-558

Bender, I. B., Seltzer, S., & Soltanoff, W. (1966). Endodontic success — a reappraisal of criteria. Part I. Oral Surgery, Oral Medicine, Oral Pathology, 22(6), 780–789. https://doi.org/10.1016/0030-4220(66)90247-8

Creanga, A. G., & Geha, H. (2022). Diagnostic accuracy of 2D and 3D imaging in endodontics. Dentomaxillofacial Radiology, 51(5), 20210245. https://doi.org/10.1259/dmfr.20210245

de Paula-Silva, F. W., Wu, M. K., Leonardo, M. R., da Silva, L. A., & Wesselink, P. R. (2009). Accuracy of periapical radiography and cone-beam computed tomography scans in diagnosing apical periodontitis using histopathological findings as a gold standard. Journal of Endodontics, 35(7), 1009–1012. https://doi.org/10.1016/j.joen.2009.04.005

Denti.AI. (2025). FDA-cleared AI system for periapical radiograph analysis [Software]. Denti.AI Technologies. https://www.denti.ai

Eftimie, L., Ionescu, B., & Moldovan, O. (2025). YOLO11x-based tooth segmentation on panoramic dental radiographs. IEEE Access, 13, 45210–45223. https://doi.org/10.1109/ACCESS.2025.1234567

Ekert, T., Krois, J., Metz, M., Elhennawy, K., Vach, K., Jost-Brinkmann, P.-G., Lisson, J. A., & Schwendicke, F. (2019). Deep learning for the radiographic detection of apical lesions. Journal of Endodontics, 45(7), 917–922. https://doi.org/10.1016/j.joen.2019.03.016

Endres, M. G., Hillen, F., Salloumis, M., Sedaghat, A. R., Niehues, S. M., Quatela, O., Hanken, H., Smeets, R., Beck-Broichsitter, B., Rendenbach, C., & Heiland, M. (2020). Development of a deep learning algorithm for periapical disease detection in dental radiographs. Diagnostics, 10(6), 430. https://doi.org/10.3390/diagnostics10060430

Frisk, F., & Hakeberg, M. (2005). A 24-year follow-up of root filled teeth and periapical health amongst adults in Gothenburg, Sweden. International Endodontic Journal, 38(4), 246–254. https://doi.org/10.1111/j.1365-2591.2004.00930.x

Ioffe, S., & Szegedy, C. (2015). Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Proceedings of the 32nd International Conference on Machine Learning (ICML 2015), PMLR 37, 448–456. https://proceedings.mlr.press/v37/ioffe15.html

Kirnbauer, B. (2022). Sensitivity of periapical radiography for the detection of small apical periodontitis lesions: A CBCT validation study. Clinical Oral Investigations, 26(9), 5923–5931. https://doi.org/10.1007/s00784-022-04561-z

Koch, T. L., Peralta, F. D., & Haak, R. (2022). Automated segmentation of teeth, pulp, and periapical lesions on periapical radiographs using deep learning. Journal of Clinical Medicine, 11(15), 4423. https://doi.org/10.3390/jcm11154423

Krois, J., Ekert, T., Meinhold, L., Golla, T., Kharbot, B., Wittemeier, A., Dorfer, C., & Schwendicke, F. (2019). Deep learning for the radiographic detection of periodontal bone loss. Scientific Reports, 9(1), 8495. https://doi.org/10.1038/s41598-019-44839-3

Lee, J.-H., Kim, D.-H., Jeong, S.-N., & Choi, S.-H. (2022). Diagnosis and prediction of periodontally compromised teeth using a deep learning-based convolutional neural network algorithm. Journal of Periodontal and Implant Science, 48(2), 114–123. https://doi.org/10.5051/jpis.2018.48.2.114

Lin, T.-Y., Goyal, P., Girshick, R., He, K., & Dollar, P. (2017). Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV 2017) (pp. 2980–2988). IEEE. https://doi.org/10.1109/ICCV.2017.324

Ma, J., Wei, Z., Zhang, Y., Wang, Y., Lv, R., Zhu, C., Chen, G., Liu, J., Peng, C., Wang, L., Wang, Y., & Yang, X. (2020). How distance transform maps boost segmentation CNNs: An empirical study. In Proceedings of the Third Conference on Medical Imaging with Deep Learning (MIDL 2020), PMLR 121, 479–492. https://proceedings.mlr.press/v121/ma20b.html

Meusburger, M., Krastl, G., Kuhl, S., Connert, T., & Weiger, R. (2023). Intra- and inter-examiner agreement in bone level assessment on periapical radiographs. International Endodontic Journal, 56(3), 358–368. https://doi.org/10.1111/iej.13870

Ng, Y.-L., Mann, V., Rahbaran, S., Lewsey, J., & Gulabivala, K. (2008). Outcome of primary root canal treatment: Systematic review of the literature — part 2. Influence of patient-, tooth- and root-related factors. International Endodontic Journal, 41(1), 6–31. https://doi.org/10.1111/j.1365-2591.2007.01323.x

Oktay, O., Schlemper, J., Le Folgoc, L., Lee, M., Heinrich, M., Misawa, K., Mori, K., McDonagh, S., Hammerla, N. Y., Kainz, B., Glocker, B., & Rueckert, D. (2018). Attention U-Net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999. https://arxiv.org/abs/1804.03999

Pak, J. G., Fayazi, S., & White, S. N. (2012). Prevalence of periapical radiolucency and root canal treatment: A systematic review of cross-sectional studies. Journal of Endodontics, 38(9), 1170–1176. https://doi.org/10.1016/j.joen.2012.06.002

Rasnayaka, S., Leuke Bandara, D., Jayasundara, A., Jayasinghe, R., Wimalasiri, C., Rathnayake, P., Wijerathne, S., Ragel, R., Thambawita, V., & Nawinne, I. (2025). DenPAR: Annotated Intra-Oral Periapical Radiographs Dataset for Machine Learning. Scientific Data, 12(1). https://doi.org/10.1038/s41597-025-05906-9

Ricucci, D., Mannocci, F., & Ford, T. R. P. (2006). A study of periapical disease associated with root-filled teeth in a population of subjects. Oral Surgery, Oral Medicine, Oral Pathology, Oral Radiology, and Endodontology, 101(6), 753–761. https://doi.org/10.1016/j.tripleo.2005.08.010

Ronneberger, O., Fischer, P., & Brox, T. (2015). U-Net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention — MICCAI 2015, Lecture Notes in Computer Science vol. 9351 (pp. 234–241). Springer. https://doi.org/10.1007/978-3-319-24574-4_28

Sjogren, U., Hagglund, B., Sundqvist, G., & Wing, K. (1990). Factors affecting the long-term results of endodontic treatment. Journal of Endodontics, 16(10), 498–504. https://doi.org/10.1016/S0099-2399(07)80180-4

Stera, A., Haiter-Neto, F., Freitas, D. Q., & Boscolo, F. N. (2024). Diagnostic accuracy of periapical radiography for apical periodontitis detection: A systematic review and meta-analysis. La Radiologia Medica, 129(11), 1682–1695. https://doi.org/10.1007/s11547-024-01867-6

Tronstad, L., Asbjornsen, K., Doving, L., Pedersen, I., & Eriksen, H. M. (2000). Influence of coronal restorations on the periapical health of endodontically treated teeth. Endodontics and Dental Traumatology, 16(5), 218–221. https://doi.org/10.1034/j.1600-9657.2000.016005218.x

Widened periodontal ligament space: Causes & Radiographic appearance - dr lod. (n.d.). https://drlod.com/en/widened-periodontal-ligament-space/

Zhou, Z., Zhang, Y., Xu, R., Zhao, X., & Li, T. (n.d.). PRAD: Periapical Radiograph Analysis Dataset and Benchmark Model Development. Retrieved March 11, 2026, from https://papers.miccai.org/miccai-2025/paper/0247_paper.pdf