Oral and maxillofacial radiology is essential for diagnosis, treatment planning, and personalized care, yet many tasks remain manual, time-consuming, and operator-dependent. This PhD thesis develops artificial intelligence (AI) tools for automated cephalometric landmarking from CBCT, tooth segmentation from CBCT, crown segmentation from intraoral scans (IOS), and multimodal fusion of CBCT/IOS, with the final aim to pose the basis to create patient-specific virtual heads. A multicentric annotated dataset was assembled, and deep learning models, including 3D V-Net, 3D attention-based U-Net, graph convolutional networks and registration algorithms, achieved high accuracy in all mentioned tasks: mean radial error of 1.95 ± 1.06 mm for landmarks annotation, mean Dice scores of 0.80 for multilabel tooth segmentation, overall accuracy of 0.94–0.95 (upper-lower arch) for crown segmentation, and sub-millimetre precision in CBCT/IOS fusion. AI-based pipelines were faster than manual methods, reproducible, and clinically feasible. The integrated framework supports diagnostics, treatment planning, surgical simulation, and education, establishing a scalable, modular foundation for fully automated virtual patient creation, with future expansion toward standardization, extended reality, and broader medical applications.
La radiologia orale e maxillo-facciale è essenziale per la diagnosi, la pianificazione del trattamento e la cura personalizzata, tuttavia molti compiti restano manuali, dispendiosi in termini di tempo e dipendenti dall’operatore. Questa tesi di dottorato sviluppa strumenti di intelligenza artificiale (AI) per il rilevamento automatizzato dei landmark cefalometrici da CBCT, la segmentazione dei denti da CBCT, la segmentazione delle corone da scansioni intraorali (IOS) e la fusione multimodale CBCT/IOS, con l’obiettivo finale di porre le basi per la creazione di teste virtuali paziente-specifiche. È stato assemblato un dataset multicentrico annotato, e i modelli di deep learning, tra cui 3D V-Net, 3D U-Net basata su attention, reti convoluzionali su grafi e algoritmi di registrazione, hanno raggiunto elevata accuratezza in tutti i compiti menzionati: errore radiale medio di 1.95 ± 1.06 mm per l’annotazione dei landmark, Dice score medio di 0.80 per la segmentazione multi-etichetta dei denti, accuratezza complessiva di 0.94–0.95 (arcata superiore e inferiore) per la segmentazione delle corone e precisione sub-millimetrica nella fusione CBCT/IOS. Le procedure automatizzate basate su AI sono risultate più rapide dei metodi manuali, riproducibili e clinicamente fattibili. Il sistema integrato supporta la diagnosi, la pianificazione del trattamento, la simulazione chirurgica e l’educazione, creando una base modulare e scalabile per la creazione completamente automatizzata di pazienti virtuali, con prospettive future di standardizzazione, realtà estesa e applicazioni mediche più ampie.
AI-driven methods in maxillofacial evaluation and therapy planning
Baldini, Benedetta
2025/2026
Abstract
Oral and maxillofacial radiology is essential for diagnosis, treatment planning, and personalized care, yet many tasks remain manual, time-consuming, and operator-dependent. This PhD thesis develops artificial intelligence (AI) tools for automated cephalometric landmarking from CBCT, tooth segmentation from CBCT, crown segmentation from intraoral scans (IOS), and multimodal fusion of CBCT/IOS, with the final aim to pose the basis to create patient-specific virtual heads. A multicentric annotated dataset was assembled, and deep learning models, including 3D V-Net, 3D attention-based U-Net, graph convolutional networks and registration algorithms, achieved high accuracy in all mentioned tasks: mean radial error of 1.95 ± 1.06 mm for landmarks annotation, mean Dice scores of 0.80 for multilabel tooth segmentation, overall accuracy of 0.94–0.95 (upper-lower arch) for crown segmentation, and sub-millimetre precision in CBCT/IOS fusion. AI-based pipelines were faster than manual methods, reproducible, and clinically feasible. The integrated framework supports diagnostics, treatment planning, surgical simulation, and education, establishing a scalable, modular foundation for fully automated virtual patient creation, with future expansion toward standardization, extended reality, and broader medical applications.| File | Dimensione | Formato | |
|---|---|---|---|
|
Baldini final thesis.pdf
accessibile in internet per tutti
Descrizione: Ph.D. thesis manuscript
Dimensione
4.07 MB
Formato
Adobe PDF
|
4.07 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/248937