Skip to content

MeshFL Demo

Jump to bottom
Mohamed E. Masoud edited this page Apr 3, 2026 · 1 revision

Overview

This demo and Colab version provides a complete walkthrough of decentralized training and inference using MeshFL for 3D MRI segmentation.

It demonstrates:

  • Federated (decentralized) training across multiple sites
  • Training convergence behavior (Loss and Dice)
  • Learning rate scheduling
  • Final segmentation output on MRI data

The goal is to show that MeshFL achieves stable and meaningful learning without centralizing data.

Training Configuration

  • Training rounds (epochs): 150
  • Batch size: 1
  • Multi-site decentralized setup

Note on Performance

This demo uses 150 training rounds for efficiency.

Increasing the number of rounds (e.g., 300–350) typically results in:

  • Higher Dice scores (often exceeding 0.93)
  • More stable convergence

Training Behavior

image

In decentralized training with small datasets:

  • Loss decreases steadily over time
  • Dice score increases as the model learns
  • Minor fluctuations are expected due to:
    • small batch size
    • distributed data across sites

We focus on overall trends rather than per-epoch noise.

image

The Demo plots illustrate:

  • Train Loss convergence
  • Train Dice improvement
  • Learning rate schedule

Observations:

  • Consistent loss reduction across sites
  • Dice scores improve steadily
  • Similar trends across sites indicate stable aggregation

Inference and Visualization

A trained model is applied to a sample MRI volume.

We visualize:

  • Input MRI slice
  • Predicted segmentations at different checkpoints during the training
  • Overlay of segmentation on MRI
image

Final Segmentation Result

  • The model successfully identifies anatomical regions after 150 rounds with possibility to improve with increasing further the number of rounds.
  • Segmentation boundaries align with structural features
  • Overlay confirms spatial consistency

This demonstrates that MeshFL produces meaningful segmentation in a decentralized setting.

Notes

  • Dataset is intentionally small for demonstration
  • Validation curves may appear noisy due to:
    • batch size of 1
    • limited samples per site
  • Despite this, overall trends confirm stable learning

Key Takeaways

  • MeshFL enables decentralized training without data sharing
  • Training is stable across multiple sites
  • The model achieves meaningful segmentation performance
  • Performance improves further with longer training

Clone this wiki locally