DFG MS Plexus

This repository contains the code used to investigate Choroid Plexus (ChP) imaging characteristics as potential biomarkers of multiple sclerosis (MS).

The ChP produces cerebrospinal fluid (CSF) and forms the blood–CSF barrier. Previous studies have associated ChP enlargement with neuroinflammatory disease. However, volume measurements alone may not capture clinically relevant differences in tissue shape and texture.

This project therefore combines:

• Choroid plexus segmentation using nnU-Net
• Radiomic shape and texture features extracted with PyRadiomics
• ALPS imaging metrics
• Convolutional neural network features derived from cropped MRI regions
• Late-fusion models for multiple-sclerosis classification

Quick Start

1. Create and activate a Python environment of your choice with Python 3.10 or newer.
2. Install nnUNet from here if segmentation training or inference is needed.
3. Install the pinned project requirements via pip install -r requirements.txt
4. Install the local package via pip install -e .

Usage

(Pre)processing

• You can normalize MRIs following FreeSurfer with

  python scripts/convert_freesurfer.py \
      --source_dir <...> \
      --output_dir <...> \
      --tmp_dir <...>
  

• You can extract cropped Regions of Interest (RoIs) around the ChPs with

  python scripts/crop_mris.py \
      --source_img_dir <...> \
      --source_mask_dir <...> \
      --output_img_dir <...> \
      --output_mask_dir <...> \
      --target_size <...>
  

Splits

Create reusable and consistent patient-wise train/test splits with

python scripts/make_splits.py \
    --root <...> \
    --target <...> \
    --fts_file <...> \
    --train_file <...> \
    --test_file <...> \
    --test_split_size <...> \
    --random_seed <...> \
    --cv_seeds <...>

Training nnU-Net for ChP Segmentation

Please follow the process described here to prepare your dataset for training nnU-Net.

You can then use bash scripts/nnunet_train_finetuned_folds_3d_fullres.sh to train your own model using the configuration we used.

For evaluation, first predict the test set masks as described here.

Then evaluate against the ground truth using python scripts/eval_nnunet.py

PyRadiomics Feature Extraction

Extract PyRadiomics features from segmented ChPs with

python scripts/extract_radiomics_features.py \
    --img_dir <...> \
    --mask_dir <...> \
    --pyradiomics_conf <...> \
    --target_filename <...>

notebooks/analyse___pyradiomics_features.ipynb implements feature pre-processing and analysis examples.

Feature-Based MS Classifier Training

The notebooks/train___*.ipynb notebooks implement different classifiers to predict MS from PyRadiomics features and ALPS imaging metrics.

CNN MS Classifier Training

The notebook notebooks/train___cnn_clf.ipynb specifically trains a CNN classifier on extracted RoIs around the ChP.

Combining CNN and Feature-Based Classifiers with Late Fusion

notebooks/eval___late_fusion___*.ipynb

Results

Model	Accuracy	Macro Precision	Macro Recall	Macro F1
Balanced Bagging	0.678 ± 0.023	0.486 ± 0.020	0.533 ± 0.018	0.506 ± 0.017
Balanced Random Forest	0.609 ± 0.032	0.493 ± 0.023	0.519 ± 0.022	0.501 ± 0.024
CatBoost	0.630 ± 0.036	0.527 ± 0.051	0.565 ± 0.056	0.536 ± 0.053
LightGBM	0.672 ± 0.028	0.552 ± 0.065	0.545 ± 0.031	0.545 ± 0.047
XGBoost	0.678 ± 0.008	0.547 ± 0.022	0.575 ± 0.010	0.556 ± 0.018
Logistic Regression	0.564 ± 0.031	0.568 ± 0.006	0.532 ± 0.017	0.530 ± 0.015
MLP	0.684 ± 0.039	0.560 ± 0.042	0.544 ± 0.023	0.547 ± 0.026
TabNet	0.654 ± 0.074	0.545 ± 0.077	0.542 ± 0.073	0.539 ± 0.072
TabPFN	0.561 ± 0.174	0.524 ± 0.020	0.536 ± 0.024	0.484 ± 0.068
CNN (ChP RoI)	0.800 ± 0.039	0.675 ± 0.050	0.646 ± 0.084	0.647 ± 0.061
LightGBM + CNN (Late Fusion)	0.866 ± 0.038	0.857 ± 0.093	0.726 ± 0.053	0.745 ± 0.058
XGBoost + CNN (Late Fusion)	0.848 ± 0.045	0.787 ± 0.057	0.741 ± 0.044	0.749 ± 0.051
MLP + CNN (Late Fusion)	0.830 ± 0.039	0.742 ± 0.101	0.712 ± 0.080	0.718 ± 0.085

Values are reported as mean ± standard deviation across five random seeds. Italic values indicate the best feature-based scores and bold values the best overall scores per metric.

Combining both model families with Late Fusion effectively fuses the feature-based models' sharper, more polarized probabilites with the CNN's softer distributions. This mechanism yields our strongest predictor with an F1-score of 0.741 ± 0.044 in the case of XGBoost + CNN, demonstrating an absolute increase of roughly 10% over the CNN and 19% over XGBoost alone.

License

This project is licensed under the Apache License 2.0. See the LICENSE file for details.

Acknowledgments

This work was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – 515302522 / SPP 2177. The funders had no role in method design, data selection and analysis, decision to publish, or preparation of the corresponding manuscript.