NA-MIC Project Weeks

Back to Projects List

Fine-Tuning SimCortex on Expert-Annotated Cortical Surfaces for Enhanced Topological Accuracy

Key Investigators

• Kaveh Moradkhani (École de Technologie Supérieure, Canada)
• Sylvain Bouix (École de Technologie Supérieure, Canada)
• Jarrett Rushmore (Boston University Medical, USA)

Project Description

SimCortex is a deep-learning framework that reconstructs all four cortical surfaces (left/right white matter and pial) from T1-weighted MRI, with a focus on minimizing inter-surface collisions and self-intersections while maintaining high geometric fidelity. To improve robustness and generalization, we fine-tune SimCortex—originally trained on FreeSurfer-generated segmentations—using a set of 50 expert-annotated MRI volumes.

---

Objectives

1. Fine‐tune SimCortex with high‐quality, manually labeled data
   Fine-tune the pre‐trained SimCortex model using 50 expert‐annotated MRI segmentations to improve anatomical accuracy and reduce geometric artifacts.

2. Compare fine‐tuned variants against the baseline
   Evaluate several fine‐tuned configurations using geometric metrics (Chamfer Distance, ASSD, HD) and topological consistency (SIF), and compare them to the baseline SimCortex model.

3. Visually validate reconstructions in 3D Slicer
   Load predicted cortical surfaces into 3D Slicer and assess anatomical plausibility with expert guidance.

---

Approach and Plan

1. Preprocessing: Convert expert-segmented MRI volumes into a format compatible with SimCortex.
2. Model Fine-Tuning: Use manual segmentations to fine-tune the original SimCortex model.
3. Evaluation: Measure Chamfer, ASSD, HD, and %SIF for each configuration.
4. Baseline Comparison: Compare all metrics with the baseline SimCortex model.
5. Expert Review: Collaborate with Prof. Jarrett Rushmore for visual validation.
6. Selection: Choose the best configuration based on both quantitative metrics and expert review.

---

Progress and Next Steps

We fine-tuned the pre-trained SimCortex model (SimCortex_M) using 50 high-quality, expert-annotated MRI segmentations. The model was evaluated against the original SimCortex baseline using geometric and topological metrics, including Chamfer Distance, Hausdorff Distance (HD), Average Symmetric Surface Distance (ASSD), and Self-Intersection Fraction (SIF).

Both quantitative and visual evaluations demonstrate the benefits of fine-tuning:

• SimCortex_M achieves lower errors across all evaluated metrics.
• The red surface (fine-tuned) shows better alignment with the turquoise manual ground truth, while the yellow surface (baseline) exhibits greater deviation.
• These results confirm that using expert-labeled data leads to improved anatomical and topological accuracy.

🚀 Future Direction

• As a next step, we plan to develop a more robust version of SimCortex as a 3D Slicer extension. This extension will aim to produce more accurate cortical surfaces with minimal inter-surface collisions and eliminate self-intersections, enhancing its usability for both clinical and research workflows.

Illustrations

📊 Quantitative Evaluation Results

The table below compares baseline and fine-tuned model performance across all surfaces.

🧠 Visual Comparison of Cortical Surfaces

This Figure illustrates the differences between ground truth and predictions.

• Cyan: Manual ground truth
• Red: Fine-tuned model (SimCortex_M)
• Yellow: Baseline model (SimCortex)

🧩 Full Brain View

🔍 Zoomed-In View

🧠 Reconstructed Cortical Surfaces (Fine-Tuned Model)