| Titre : | DenteDot: An AI-Based Platfrom For Automated Dental Disease Detection and Intelligent Clinical Decision Support |
| Auteurs : | Boukhorata Elhosseyn, Auteur ; Mostefai abdelkader, Directeur de thèse |
| Type de document : | texte manuscrit |
| Editeur : | Université de Saïda – Dr. Moulay Tahar – Faculté des Mathématiques, de l’Informatique et des Télécommunications, 2025/2026 |
| Format : | 62 ص |
| Accompagnement : | CD |
| Langues: | Français |
| Index. décimale : | BUC-M 008552 |
| Catégories : |
Master en informatique Spécialité : Modélisation Informatique des Connaissances et du Raisonnement |
| Résumé : |
The manual interpretation of dental radiographs in daily clinical practice is a highly demand-
ing task that often leads to visual fatigue and diagnostic inconsistencies. To address this challenge, this thesis presents DenteDot—an automated clinical decision support system de- signed to accurately segment anatomical structures and detect pathological conditions using advanced deep learning architectures. Architecturally, the system is decoupled into two specialized microservices to ensure sta- bility and modularity. The primary radiographic module, managed within app.py, processes panoramic and periapical X-rays by orchestrating three concurrent deep learning models: YOLOv11x-seg for the precise instance segmentation of all 32 teeth, DeepLabV3+ for the robust detection of deep dentin caries, and a specialized Kennedy YOLOv11s model designed to analyze the jaw for retained roots, periodontally compromised teeth, and edentulous arches, enabling automated Kennedy Class I-IV categorization. Furthermore, to address surface-level pathologies, a secondary microservice (teeth_marking_app.py) was developed specifically for intraoral photography. This dedicated FastAPI endpoint uti- lizes a 9-class instance segmentation model to delineate surface abrasions, restorations, ce- ramic crowns, and six distinct stages of caries, while demonstrating robustness against spec- ular highlights caused by saliva. The models were trained on three comprehensive datasets. The primary dataset com- prised 598 radiographs featuring over 15,300 manual polygon annotations for teeth map- ping. Severe class imbalances, notably regarding extracted third molars, were mitigated using copy-paste augmentation protocols. Additionally, the system utilized a specialized Kennedy OPG Dataset containing 1,250 radiographs for structural classification, and an Intraoral Pho- tography Dataset consisting of 1,320 high-resolution (15-megapixel) images across 295 pa- tients for the surface marking task. Finally, to ensure real-time clinical applicability, the PyTorch models were optimized and exported to the ONNX runtime, significantly reducing inference latency to the millisecond scale. |
Exemplaires
| Code-barres | Cote | Support | Localisation | Section | Disponibilité |
|---|---|---|---|---|---|
| aucun exemplaire |
Documents numériques (1)
BUC-M 008552 Adobe Acrobat PDF |
