⚡ electrical engineering

Impact of Labeling Inaccuracy and Image Noise on Tooth Segmentation in Panoramic Radiographs using Federated, Centralized and Local Learning

This study demonstrates that Federated Learning, utilizing per-client loss trajectories for anomaly detection, matches or outperforms Centralized and Local Learning in tooth segmentation across various data corruption scenarios while effectively preserving patient privacy.

Original authors: Johan Andreas Balle Rubak, Khuram Naveed, Sanyam Jain, Lukas Esterle, Alexandros Iosifidis, Ruben Pauwels

Published 2026-03-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Johan Andreas Balle Rubak, Khuram Naveed, Sanyam Jain, Lukas Esterle, Alexandros Iosifidis, Ruben Pauwels

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to spot every single tooth in a panoramic X-ray of a mouth. This is a tricky job for a computer, but it's essential for dentists to plan treatments.

The problem is that dental clinics are like isolated islands. They have their own X-ray machines and their own patients, but they can't easily share the actual patient images because of strict privacy laws (like GDPR). It's like trying to bake a world-class cake, but every baker is forbidden from sharing their secret ingredients with anyone else.

This paper explores three different ways to solve this "island problem" to train the best possible AI tooth-spotter.

The Three Teams

The Local Learner (The Solo Baker):
Imagine a baker who only has flour from their own pantry. They try to bake a cake using only their local ingredients.
- The Result: The cake is okay, but it's not very good because they don't have enough variety. If their flour is slightly bad, the whole cake fails.
The Centralized Learner (The Big Central Kitchen):
Imagine a giant kitchen where every baker sends all their flour, eggs, and sugar to one central location. One master chef mixes it all together to bake the perfect cake.
- The Result: This usually makes the best cake because it has the most ingredients.
- The Catch: In the real world, bakeries (hospitals) can't send their ingredients (patient data) to a central kitchen because of privacy laws. It's illegal to ship the flour.
The Federated Learner (The Secret Recipe Exchange):
This is the star of the show. Instead of sending the ingredients, every baker sends a note describing how they mixed their batter. A central coordinator collects all these notes, mixes them together to create a "super-recipe," and sends the new recipe back to every baker.
- The Result: The bakers never leave their kitchens, and no one sees the actual ingredients, but they all learn from each other to bake a better cake.

The Experiment: What Happens When Things Go Wrong?

The researchers wanted to see if this "Secret Recipe Exchange" (Federated Learning) could handle real-world messiness. They set up four scenarios:

Scenario 1: The Perfect Day (Baseline)
Everyone has good ingredients and follows the recipe perfectly.
- Outcome: The Federated team baked a cake just as good as the Central Kitchen, and much better than the Solo Bakers.
Scenario 2: The Messy Note-Taker (Label Inaccuracy)
One baker (Client 0) is sloppy. Instead of tracing the teeth perfectly, they draw the outlines too wide or forget to draw a tooth entirely.
- Outcome: The Federated team was surprisingly resilient. Even though one baker was sending bad notes, the group ignored the bad advice enough to still bake a great cake. The Central Kitchen also did well, but the Solo Baker who was sloppy made a terrible cake.
Scenario 3: The Foggy Camera (Image Noise)
One baker's camera is dirty, adding static or "snow" to the photos.
- Outcome: Again, the Federated team handled it well. They learned from the other 4 clean bakers and didn't let the one foggy camera ruin the whole group's recipe.
Scenario 4: The "Bad Apple" Removal
The researchers realized one baker was consistently sending bad notes. They kicked that baker out of the group.
- Outcome: The Federated group got even better.

The Secret Superpower: The "Loss Curve" Fingerprint

Here is the coolest part of the paper. The researchers found a way to spot the "bad baker" automatically without even looking at the final cake.

They watched the training loss (a score that measures how confused the AI is).

The good bakers' confusion scores went down smoothly.
The bad baker's confusion score went up or got stuck.

It's like a teacher listening to students read aloud. If one student is stumbling over every word while the others are reading smoothly, the teacher knows immediately who needs help (or who needs to be temporarily removed from the group). The paper shows that by watching these "confusion scores," the system can automatically detect and exclude the faulty client, making the whole group smarter.

The Bottom Line

Federated Learning works: It allows hospitals to collaborate on AI without breaking privacy laws.
It's tough: It can handle messy data and bad labels better than a single hospital working alone.
It's smart: It can even "fire" the bad data sources automatically by watching their behavior.

In short, this paper proves that we can build a super-smart dental AI by having clinics work together secretly, rather than forcing them to share their private patient files. It's a win for privacy and a win for better healthcare.

1. Problem Statement

The integration of Artificial Intelligence (AI) into dental diagnostics is hindered by data privacy regulations (e.g., GDPR), ethical concerns, and the logistical difficulty of aggregating sensitive medical data from multiple institutions. While Centralized Learning (CL) offers high performance by pooling data, it violates privacy constraints. Local Learning (LL) preserves privacy but suffers from limited data diversity and poor generalization. Federated Learning (FL) is proposed as a solution to train models collaboratively without sharing raw data.

However, real-world FL deployment faces specific challenges:

Data Heterogeneity: Variations in imaging equipment and patient demographics across institutions.
Labeling Inconsistency: Human annotators may produce noisy, inaccurate, or incomplete ground-truth masks.
Image Quality Degradation: Noise or artifacts in specific client datasets.
Faulty Clients: The risk of a single client with corrupted data negatively impacting the global model.

This study investigates whether FL can maintain robustness against these specific data corruptions while matching or exceeding the performance of CL and LL in tooth segmentation tasks.

2. Methodology

Dataset and Preprocessing

Data Source: The study utilized the open-source "A dual labeled dataset" containing 2,066 panoramic radiographs from six institutions (three hospitals and three clinics) in China and Denmark.
Annotations: Teeth were delineated using polygonal annotations by four dentists, labeled according to FDI notation (33 distinct labels).
Preprocessing: Images were resized to $512 \times 1088$ pixels (preserving aspect ratio) and padded/cropped as necessary.
Split: 10% held out as a centralized test set; the remaining data was partitioned for training and validation.

Model Architecture

Architecture: Attention U-Net. This was chosen over standard U-Net for its ability to suppress irrelevant regions and focus on critical anatomical structures via attention gates.
Loss Function: Dice Loss.
Optimizer: AdamW ( $lr = 1 \times 10^{-4}$ ).
Training: 50 epochs total. For FL, these were distributed over 5 rounds (10 epochs per round).

Learning Paradigms

The study compared three paradigms:

Local Learning (LL): Training on a single client's data.
Centralized Learning (CL): Training on all pooled data.
Federated Learning (FL): Training via the Flower AI framework using FedAvg (Federated Averaging). Clients train locally and share model weights, not data.

Experimental Configurations

Four scenarios were tested to simulate real-world imperfections:

Baseline: Unaltered data.
Label Manipulation: Simulated poor annotation by one client (Client 0). This involved mask dilation (random kernels 3–11) and 10% random omission of tooth masks.
Image Quality Manipulation: Simulated noise for Client 0 by adding Gaussian noise ( $\mu=0, \sigma=25$ ).
Faulty Client Exclusion: Client 0 was excluded from training in CL and FL to test the impact of removing bad data.

Evaluation Metrics

Primary: Dice Score (Sørensen–Dice coefficient).
Secondary: Intersection over Union (IoU), Hausdorff Distance (HD), 95th percentile HD (HD95), and Average Symmetric Surface Distance (ASSD).
Statistical Analysis: Wilcoxon signed-rank test with Bonferroni correction ( $\alpha_{corr} \approx 0.00278$ ) to determine significance.

3. Key Results

Performance Comparison (Baseline)

FL achieved the highest median Dice score (0.94889), slightly outperforming CL (0.94706) and significantly outperforming LL (0.93557–0.94026).
FL and CL showed comparable performance in most metrics, while LL consistently underperformed due to limited data diversity.

Robustness to Corruption

Label Manipulation: FL maintained the best performance (Dice: 0.94884) despite Client 0's corrupted labels. CL dropped to 0.94183, and LL (specifically the model trained on the corrupted client, LL0) dropped significantly to 0.93003.
Image Noise: FL remained the leader (Dice: 0.94853), closely followed by CL (0.94787). LL0 suffered the most (Dice: 0.93179).
Faulty Client Exclusion: When the corrupted client was removed, FL (Dice: 0.94790) slightly outperformed CL (Dice: 0.94550), suggesting FL benefits from the collaborative nature even with fewer participants.

Anomaly Detection via Loss Curves

A critical finding was that per-client training loss trajectories served as an effective anomaly detector.
The "faulty" client (Client 0) exhibited elevated training loss and slower convergence compared to other clients.
Interestingly, this elevated loss did not appear in the validation loss (which remained similar to other clients), likely due to Batch Normalization statistics behaving differently during training vs. validation.
This divergence allows for the automatic identification and exclusion of faulty clients before they degrade the global model.

Visual Analysis

FL vs. CL: Segmentation results were visually very similar.
Error Patterns: FL errors tended to be spatially clustered (large contiguous false negatives/positives), whereas CL errors were more scattered. FL showed slightly higher sensitivity to outlier cases (e.g., sparse dentition, implants).

4. Key Contributions

Empirical Validation of FL Robustness: The study demonstrates that FL not only matches but often exceeds CL performance in dental segmentation, even when specific clients contribute noisy labels or degraded images.
Anomaly Detection Mechanism: The authors propose a practical method for client-level anomaly detection by monitoring training loss curves. This allows systems to flag and exclude "faulty" participants without needing to inspect raw data, preserving privacy.
Comprehensive Benchmarking: It provides a rigorous comparison of FL, CL, and LL under controlled, realistic corruption scenarios (label noise, image noise, and client exclusion), filling a gap in dental AI literature.
Implementation Framework: The successful implementation using Flower AI and FedAvg on a multi-center dataset serves as a blueprint for scalable, privacy-preserving clinical AI deployment.

5. Significance and Conclusion

This paper validates Federated Learning as a viable, privacy-preserving alternative to centralized data aggregation for medical imaging. The findings suggest that:

Privacy vs. Performance: FL can achieve performance comparable to CL without compromising patient data privacy.
Resilience: FL models are more resilient to data heterogeneity and labeling errors than Local Learning models.
Operational Feasibility: The ability to detect and exclude faulty clients via loss monitoring makes FL systems more robust for real-world clinical deployment where data quality varies.

The study concludes that FL is a practical approach for scalable clinical AI, offering a balance between high-performance model training and strict adherence to data privacy regulations, provided that mechanisms for monitoring client contribution quality are in place.