Edit this page

NA-MIC Project Weeks

Back to Projects List

Fine-tuning SimCortex Using Manually Corrected Cortical Annotations

Key Investigators

Project Description

SimCortex v2 is a deep learning pipeline for cortical surface reconstruction from brain MRI. In this project, we will fine-tune the existing SimCortex v2 model using manually corrected segmentations and cortical surfaces.

The goal is to evaluate whether manual supervision can improve the reconstructed white and pial surfaces compared with the current SimCortex v2 baseline.

As a practical outcome, we will also prepare a 3D Slicer extension for SimCortex. The extension runs SimCortex through Docker from a native T1-weighted MRI and loads the reconstructed cortical surfaces back into Slicer for visualization.

Objective

  1. Fine-tune SimCortex v2 using manually corrected segmentations and cortical surfaces, and evaluate the fine-tuned model on held-out test data.

  2. Prepare the SimCortex 3D Slicer extension for public release, so users can run the pipeline and visualize the reconstructed surfaces directly in 3D Slicer.

Approach and Plan

  1. Review the manually corrected segmentations and cortical surfaces.
  2. Prepare the manual annotations in a format compatible with the SimCortex v2 training pipeline.
  3. Fine-tune the relevant SimCortex v2 stages using the manual annotations.
  4. Run inference with both the baseline and fine-tuned models on the same test data.
  5. Compare the results using surface metrics, topology-related checks, and visual inspection.
  6. Test the local SimCortex 3D Slicer extension and prepare the repository for public use.

Progress and Next Steps

  1. SimCortex v2 is already available as an open-source cortical surface reconstruction pipeline.
  2. Manually corrected segmentations and cortical surfaces are available for fine-tuning.
  3. A local prototype of the SimCortex 3D Slicer extension has been developed.
  4. The extension can run SimCortex through Docker from a T1-weighted MRI and load the reconstructed white and pial surfaces back into Slicer.
  5. Next steps are to finalize the fine-tuning workflow, run initial experiments, evaluate the results, and prepare the Slicer extension for public release.

Illustrations

The main illustration shows the SimCortex pipeline from brain MRI to segmentation, initial surfaces, deformation, and final predicted cortical surfaces.

Image

A second illustration shows the local SimCortex 3D Slicer extension with reconstructed white and pial surfaces loaded in Slicer.

Image

Results and Outputs from Project Week

During Project Week, we worked on two main directions: model fine-tuning and 3D Slicer integration.

For the fine-tuning part, we trained SimCortex using 30 manually corrected segmentations and cortical surfaces. We tested different training configurations and evaluated the models on three independent datasets (CNP, ds001486, HCP/OASIS), covering 120 test subjects in total. The evaluation includes surface distance metrics (Chamfer, ASSD, HD90), self-intersection fraction (SIF), cortical thickness error, and inter-surface collision checks.

The tables below summarize the quantitative comparison between the baseline SimCortex and SimCortex FineTune.

image

image

For the 3D Slicer part, the SimCortex extension was merged into the official 3D Slicer Extensions Index. Users can now install it directly from Slicer, run the full pipeline from a T1-weighted MRI through Docker, and load the reconstructed white and pial surfaces back into Slicer for visualization.

Background and References