Deep Learning-based Assessment of the Relation Between the Third Molar and Mandibular Canal on Panoramic Radiographs using Local, Centralized, and Federated Learning

This study demonstrates that while centralized learning yields the highest accuracy in classifying third molar and mandibular canal overlap on panoramic radiographs, federated learning serves as a superior privacy-preserving alternative that significantly outperforms local learning models.

The Gist

Here is an explanation of the paper, translated into everyday language with some creative analogies.

The Big Picture: The "Wisdom Tooth" Problem

Imagine you have a wisdom tooth (the third molar) that is stuck in your jaw. Right underneath it runs a tiny, fragile nerve (the mandibular canal). If a dentist pulls that tooth out, they need to be very careful not to nick that nerve, or the patient could lose feeling in their lip or chin.

To check the risk, dentists usually take a flat X-ray (a panoramic radiograph). But looking at a flat 2D picture to guess where a 3D nerve is hidden is tricky. Sometimes the tooth looks like it's touching the nerve, but it's actually just "in front" of it.

The Goal: The researchers wanted to build an AI (a computer program) that can look at these X-rays and instantly say: "Safe to pull" or "Danger! This tooth is touching the nerve."

The Challenge: The "Data Privacy" Dilemma

To teach an AI to do this well, you need a lot of examples. But here's the catch: Dental X-rays contain private patient data. Hospitals and clinics cannot just email thousands of X-rays to a central server to train the AI because of privacy laws.

So, the researchers asked: How do we train a smart AI without moving the private data?

They tested three different "training schools" to see which one produced the best teacher (AI model).

---

The Three Training Schools

1. Local Learning (The "Isolated Student")

• The Analogy: Imagine eight different students, each sitting in their own separate room. Each student only has a small stack of X-rays from their own local clinic. They study their own stack, take a test, and get a grade.
• The Result: Each student becomes an expert on their specific stack of X-rays. But if you give them a test using X-rays from a different clinic, they get confused. They are too specialized and can't generalize.
• The Paper's Finding: These models worked okay in their own "rooms" but failed miserably when tested on new data from other places.

2. Centralized Learning (The "All-Hands Meeting")

• The Analogy: Imagine all eight students are allowed to bring their X-rays to one big conference room. They dump all the pictures on one giant table. They study the entire collection together as one team.
• The Result: This is the "Gold Standard." Because they saw every possible type of X-ray, they became the smartest, most well-rounded experts.
• The Catch: In the real world, this is often illegal or impossible because of privacy laws. You can't just move all the patient data to one place.

3. Federated Learning (The "Secret Messenger")

• The Analogy: This is the clever middle ground. The students stay in their own rooms (data stays private). However, they all have a "brain" (the AI model).
  • The teacher sends a "starter brain" to every student.
  • Each student studies their own local X-rays and updates their brain.
  • They send only the updates (the "lessons learned") back to the teacher, not the X-rays themselves.
  • The teacher mixes all the updates together to create a new, smarter "Global Brain" and sends it back out.
• The Result: They never saw each other's data, but they learned from each other's experiences.
• The Paper's Finding: This method didn't quite reach the genius level of the "All-Hands Meeting," but it was much better than the isolated students. It was a great privacy-friendly compromise.

---

The Key Takeaways

1. The "Privacy vs. Performance" Trade-off
The study found that if you can break privacy rules and pool all the data (Centralized Learning), you get the best AI. However, if you must keep data private (Federated Learning), you still get a very good AI—much better than trying to do it alone.

2. The "Overfitting" Trap
The researchers noticed that the "Isolated Students" (Local Learning) were actually "cheating" a bit. They memorized the specific lighting, camera angles, or quirks of their own local clinic's X-rays. They looked smart on their own tests but failed when the test changed slightly. This is called overfitting. The Federated Learning method helped stop this by forcing the AI to learn general rules that work everywhere.

3. The "Teacher's Eye" (Grad-CAM)
The researchers used a special tool called Grad-CAM to see what the AI was looking at.

• The Good AI (Centralized/Federated): When making a decision, it looked right at the tooth and the nerve. It focused on the anatomy.
• The Bad AI (Local): It often looked at random spots, like the edge of the image or the background noise. It was guessing based on the wrong clues.

The Bottom Line for Dentists

If you want to build an AI to help triage wisdom teeth:

• Best Case: Pool all your data (if legal) to get the smartest model.
• Realistic Case: Use Federated Learning. It allows clinics to collaborate and build a smart, privacy-safe tool that is far superior to any single clinic trying to do it alone.
• Worst Case: Don't just train a model on one clinic's data and expect it to work everywhere; it will likely fail.

In short: Federated Learning is the "Swiss Army Knife" of medical AI—it's not quite as powerful as the full toolbox (Centralized), but it's infinitely better than working with just a single screwdriver (Local).

Technical Summary

Here is a detailed technical summary of the paper "Deep Learning-based Assessment of the Relation Between the Third Molar and Mandibular Canal on Panoramic Radiographs using Local, Centralized, and Federated Learning."

1. Problem Statement

The impaction of mandibular third molars in close proximity to the mandibular canal poses a significant risk of inferior alveolar nerve injury during extraction. While Cone-Beam Computed Tomography (CBCT) offers superior 3D detail, panoramic radiographs (PR) remain the standard initial screening tool due to lower cost and radiation exposure.

• Clinical Challenge: Accurately assessing the relationship between the tooth root and the canal (3M-MC) on 2D PRs is difficult and prone to inter-observer variability. Automated deep learning (DL) tools could assist in triage and reduce unnecessary CBCT referrals.
• Data Challenge: Developing robust DL models requires large, diverse, multi-institutional datasets. However, privacy regulations and data governance often prevent the centralization of medical imaging data.
• Research Gap: There is a need to empirically compare Federated Learning (FL) against Centralized Learning (CL) and Local Learning (LL) in a realistic, non-IID (non-Independent and Identically Distributed) medical setting to determine if FL can achieve performance comparable to CL without compromising patient privacy.

2. Methodology

2.1. Dataset and Labeling

• Data Source: The study utilized panoramic radiographs from five public datasets (ADLD, Dentex, TSXK, Tufts, USPFORP).
• Region of Interest (ROI): An automated process cropped the lower third molars based on existing tooth annotations.
• Labeling Protocol: Eight professionally trained dentists (clients) annotated the images. The task was a binary classification: No-overlap vs. Overlap (merging various overlap subclasses).
• Data Distribution: The data was partitioned by labeler (client), creating a natural non-IID scenario where each client had unique class distributions, imaging devices, and population demographics.
  • Total Samples: 5,416 images (3,561 No-overlap, 1,855 Overlap).
  • Class Imbalance: Significant variation existed across clients (Overlap % ranged from 12.3% to 40.6%).

2.2. Model Architecture

• Backbone: Pre-trained ResNet-34 (ImageNet weights).
• Head: Replaced with a two-unit linear layer for binary classification.
• Preprocessing: Contrast Limited Adaptive Histogram Equalization (CLAHE), resizing to 224x224, and normalization.
• Augmentation: To address class imbalance, the minority class was upsampled via random augmentation (flips, rotations, jitter, blur) with periodic regeneration to introduce stochasticity.

2.3. Learning Paradigms Compared

The study compared three distinct training strategies using identical hyperparameters (AdamW optimizer, Binary Cross-Entropy loss, batch size 32):

1. Local Learning (LL): Each of the 8 clients trained an independent model solely on their local data.
2. Centralized Learning (CL): All data was pooled; a single model was trained on the aggregated dataset (the "gold standard" for performance).
3. Federated Learning (FL): A collaborative approach using FedAvg. Clients trained locally for 2 epochs per round; model weights were sent to a central server, aggregated (sample-weighted average), and redistributed. The process ran for 5 rounds.

2.4. Evaluation Metrics

• Metrics: Accuracy, Sensitivity, Specificity, F1-score, Balanced Accuracy, Youden's J, and AUC-ROC.
• Evaluation Strategy:
  • Local Validation: Models evaluated on each client's local validation set using client-specific optimized thresholds.
  • Centralized Test: All models evaluated on a held-out, pooled test set (10% of total data) using a single global threshold to assess generalization.
• Explainability: Grad-CAM visualizations were used to analyze which image regions the models focused on.

3. Key Results

3.1. Performance Comparison (Centralized Test Set)

• Centralized Learning (CL): Achieved the highest performance.
  • AUC: 0.831
  • Accuracy: ~0.782
• Federated Learning (FL): Showed intermediate performance, significantly outperforming LL but falling short of CL.
  • AUC: 0.757
  • Accuracy: ~0.703
• Local Learning (LL): Generalized poorly across clients.
  • AUC Range: 0.619 – 0.734 (Mean ≈ 0.672).
  • Observation: While LL models performed reasonably well on their own local data, they failed to generalize when applied to the pooled test set, exhibiting severe calibration issues (e.g., extreme false positive/negative rates under a global threshold).

3.2. Statistical Significance

• CL vs. FL: CL significantly outperformed FL (DeLong's test, $p < 0.001$).
• FL vs. LL: FL significantly outperformed 7 out of 8 LL models.
• CL vs. LL: CL significantly outperformed all individual LL models.

3.3. Training Dynamics & Overfitting

• Overfitting: All paradigms showed signs of overfitting (training loss decreased while validation loss fluctuated or increased). This was attributed to the difficulty of the task (2D projection of 3D anatomy) and the use of augmented minority class samples which may have inflated training performance without improving true generalization.
• FL Stability: FL validation loss showed fluctuations after aggregation rounds, indicative of the challenges posed by non-IID data distributions.

3.4. Explainability (Grad-CAM)

• CL & FL: Heatmaps showed focused, anatomically consistent attention on the third molar and mandibular canal regions.
• LL: Lower-performing LL models exhibited diffuse, inconsistent, or spatially scattered activations, suggesting they relied on site-specific artifacts or noise rather than robust anatomical features.

4. Key Contributions

1. Empirical Comparison: Provides a rigorous, controlled comparison of LL, CL, and FL in a realistic dental imaging scenario with 8 independent labelers and heterogeneous data sources.
2. Quantification of the "Federated Gap": Demonstrates that while FL improves robustness over local training, a performance gap remains compared to centralized learning in non-IID settings, primarily due to labeler bias and data heterogeneity.
3. Calibration Analysis: Highlights that local models often suffer from severe score miscalibration when applied globally, making a single decision rule unreliable for LL but manageable for CL and FL.
4. Server-Side Monitoring: Proposes the use of non-sensitive server-side signals (e.g., update magnitudes, local loss trajectories) to detect outliers and heterogeneity without violating privacy.
5. Clinical Workflow Insight: Validates that FL is a viable privacy-preserving alternative for multi-center collaboration, offering better generalization than isolated local models, though it requires careful thresholding and calibration strategies.

5. Significance and Conclusion

This study confirms that Centralized Learning remains the performance benchmark for automated 3M-MC assessment when data pooling is feasible. However, Federated Learning emerges as a robust, privacy-preserving middle ground that significantly outperforms isolated Local Learning models.

The findings suggest that for clinical deployment in multi-center settings:

• FL is preferred over LL to avoid the "silos" of local overfitting and poor cross-site generalization.
• Calibration is critical: A single global threshold may not work for all local models; strategies like per-site threshold optimization or sending summary metrics (e.g., Youden's J curves) to the server are recommended.
• Future Directions: To close the gap between FL and CL, future work should explore heterogeneity-aware aggregation algorithms (e.g., FedProx, personalized FL) and advanced calibration techniques.

The study underscores that while FL does not yet fully match the performance of centralized training in highly heterogeneous medical datasets, it is a necessary and effective step toward scalable, privacy-compliant AI in dentistry.