1. Introduction AI is already woven into everyday life (voice assistants, smart homes), but error cases show the critical need for robust, well‑balanced training data.
This session explores how AI and “big data” can transform diagnostics, planning, surgery, and long‑term implant care.
2. Prof. Andreas Dengel – Foundations of AI & “Good Data” End‑to‑End Learning
Traditional pipelines (feature engineering → model selection) are being supplanted by deep neural nets that learn features automatically, often extracting billions of parameters.
Data Quality & Quantity
AI performance hinges on both: too little data → poor generalization; too many noisy or unbalanced examples → spurious results.
Case Study: Label‑Free Cell Imaging
Public datasets of contrast‐microscopy images suffer from low contrast, overlapping cells, partial annotations.
Solution: generate 1.6 million fully annotated synthetic images (Nature‑published), enabling automated cell‐tracking pipelines.
Explainable AI (XAI)
Black‑box CNNs in medicine (e.g. melanoma classification) demand transparency.
Three‑step “concept‑based” approach:
Pre‑train on large general dataset (e.g. ImageNet).
Train mid‑level “concept” classifiers (dots, stripes…) on a smaller, domain‑relevant set.
Fine‑tune on target domain (skin lesions), yielding both diagnosis and human‑readable heatmaps/textual explanations.
Data Sharing & Governance
European Health Data Space aims to standardize access, interoperability, privacy—but legal, ethical, and technical hurdles remain.
3. Prof. Tim Joda – Clinical AI in Implant Workflows Focusing on four pillars of the digital implant workflow:
Diagnostics & Treatment Planning
2D & 3D segmentation: automatic tooth/implant/nerve/airway detection on panoramic, periapical, and CBCT.
Automated measurements: bone height/width, soft‑tissue biotype, emergence profiles.
Future goal: AI‑driven implant‑type, diameter, and position recommendations (still limited by lack of labeled planning datasets).
Guided & Robotic Surgery
Static guides: flapless but inflexible once fabricated.
Dynamic navigation: haptic‑feedback, real‑time 3D guidance—offers freedom to adjust intra‑operatively.
Robotic assistance: first dental‑robot cases (up to 4 mm deviation) show promise but need accuracy improvements.
Prosthetics & Virtual Patients
AI for automated landmark detection and maxillomandibular registration.
“4D simulation”: combining intraoral scans, facial scans, CBCT to preview function, esthetics, and lip support in full‑arch cases.
Remote Monitoring & Personalized Care
Biosensors embedded in the mouth to track salivary biomarkers.
Vision of a “digital twin” for each patient enabling Predictive, Preventive, Personalized, Participatory (4P) implant care.
Early deep‑learning risk models (AUC≈0.90) for implant failure show potential but require cautious validation.
4. Prof. Arjan Vissink – Harnessing “Big Data” in Implant Dentistry Current Landscape
Most “big data” studies rely on sales records, insurance claims, or retrospective cohorts—lacking detailed implant‑ and patient‑level parameters.
Registry Examples & Limitations
Implant sales (n≈97 000): 2.8 % early loss, diameter influences risk—but no data on brand, placement/loading protocols, soft‑tissue outcomes.
Surgeon volume (n≈27 000 implants): clear learning curve effect, but indications and case complexity uncontrolled.
Elderly health (n≈170 000 – 260 000): insurance data link dentures/implants to general health and costs, yet lack implant‑specific variables.
Sjögren’s cohort (n>15 000): uniform classification criteria enabled high‑quality implant survival and health analyses in an autoimmune disease setting.
Key Take‑Home:
Uniform, standardized data collection (baseline, procedural, and follow‑up variables) is essential to unlock true big‑data insights: trends, risk factors, and optimized treatment protocols.
Without agreed‑upon variable sets, studies remain limited to broad trends, not actionable, patient‑level decision support.
5. Interactive Discussion – Q&A Highlights Limitations of Guided/Robotic Surgery: – Static guides lack intra‑op flexibility; dynamic systems add radiation and cost; robotic accuracy still variable (max ~4 mm error).
Certifying Data Quality: – Balance datasets at collection; use explainability methods to detect outliers; employ multimodal networks combining images + metadata.
Predictive Modeling & Ethics: – AI risk scores could one day influence insurance coverage or contraindicate implants—raising questions about patient autonomy and access.
Data Governance: – European Health Data Space will mandate privacy and consent; generational and regional attitudes toward data sharing will shape adoption.
Overall Take‑Away AI and big data are poised to amplify clinicians’ expertise—streamlining diagnostics, guiding surgery, and personalizing long‑term care—but success hinges on high‑quality, standardized data, robust validation, and ethical frameworks that keep patient welfare front and center.