TBI Lesion Analysis Pipeline

A deep learning-based pipeline for analyzing traumatic brain injury (TBI) lesions from MRI scans. This tool segments 7 types of TBI lesions and quantifies their overlap with 10 brain anatomical zones. This was a proof of concept using Claude Code, where I found interesting Computed Tomography based Traumatic Brain Injury Quantification python code at https://github.com/nifm-gin/CT-TIQUA. I then modified the code to use MRI DICOM images, and trained using MR Brain Segmentation Challenge 2018 Data. Finally, I ran inference on some existing MRI images from a TRACTS study.

Next steps would be to make the inference and validation more rigourous and begin an investigation on the lesions identified. The sample reports are in the output folder.

Features

• MRI Preprocessing: Bias field correction, skull stripping, and normalization
• Multi-class Lesion Segmentation: Identifies 7 types of TBI lesions
• Brain Atlas Integration: Maps lesions to 10 anatomical brain zones
• Quantitative Analysis: Calculates lesion volumes per zone with CSV export
• Visualization Support: Generates overlay images and color lookup tables

Lesion Types Detected

1. Intraparenchymal hemorrhage
2. Subdural hematoma
3. Epidural hematoma
4. Subarachnoid hemorrhage
5. Intraventricular hemorrhage
6. Diffuse axonal injury
7. Contusion

Brain Zones

1. Frontal lobe (left/right)
2. Parietal lobe (left/right)
3. Temporal lobe (left/right)
4. Occipital lobe (left/right)
5. Cerebellum
6. Brainstem

Installation

# Clone the repository
cd tbi

# Install dependencies
pip install -r requirements.txt

Usage

Basic Usage

python run_pipeline.py /path/to/mri/data /path/to/output --subject-id patient001

Advanced Options

python run_pipeline.py /path/to/mri/data /path/to/output \
    --subject-id patient001 \
    --atlas /path/to/custom/atlas.nii.gz \
    --model /path/to/pretrained/model.pth \
    --save-intermediate \
    --save-probabilities \
    --log-level DEBUG

Command Line Arguments

• input: Path to input MRI (DICOM directory or NIfTI file)
• output: Output directory for results
• --subject-id: Subject identifier (default: "subject")
• --atlas: Path to custom brain atlas (optional)
• --model: Path to pretrained segmentation model (optional)
• --skip-preprocessing: Skip preprocessing step
• --save-intermediate: Save intermediate processing files
• --save-probabilities: Save lesion probability maps
• --log-level: Logging level (DEBUG/INFO/WARNING/ERROR)

Output Structure

output_directory/
├── preprocessing/
│   ├── preprocessed_mri.nii.gz
│   └── brain_mask.nii.gz
├── segmentation/
│   ├── lesion_segmentation.nii.gz
│   └── prob_*.nii.gz (if --save-probabilities)
├── registration/
│   └── atlas_in_subject_space.nii.gz
├── atlas/
│   └── zone_masks/
├── quantification/
│   ├── {subject_id}_lesion_zone_overlap.csv
│   ├── {subject_id}_zone_summary.csv
│   ├── {subject_id}_lesion_summary.csv
│   └── {subject_id}_report.txt
├── visualization/
│   ├── lesion_atlas_overlay.nii.gz
│   └── color_lookup_table.json
├── pipeline.log
└── {subject_id}_pipeline_summary.json

Key Output Files

Lesion-Zone Overlap CSV

Contains detailed volume measurements for each lesion type in each brain zone:

• Zone: Brain region name
• LesionType: Type of lesion
• Volume_mm3: Volume in cubic millimeters
• Volume_ml: Volume in milliliters

Report

Human-readable summary including:

• Total lesion volume
• Affected brain zones
• Lesion type breakdown
• Detailed lesion-zone combinations

Pipeline Architecture

1. Preprocessing (src/preprocess.py)

   • DICOM/NIfTI loading
   • N4 bias field correction
   • Skull stripping
   • Intensity normalization

2. Segmentation (src/segment.py)

   • 3D U-Net architecture (MONAI)
   • Multi-class prediction (8 classes: 7 lesions + background)
   • Volume calculation

3. Atlas Registration (src/atlas.py)

   • ANTsPy-based registration
   • 10-zone brain parcellation
   • Zone mask generation

4. Quantification (src/quantify.py)

   • Lesion-zone overlap calculation
   • Statistical summaries
   • CSV export

Model Training

The segmentation module includes training transforms for custom model development:

from src.segment import TBISegmenter

segmenter = TBISegmenter()
train_transforms = segmenter.create_training_transforms()

Requirements

• Python 3.8+
• CUDA-capable GPU (recommended)
• See requirements.txt for full dependencies

Notes

• Default model uses random initialization (requires training on TBI data)
• Atlas registration quality depends on preprocessing quality
• Processing time varies with image size and GPU availability

Citation

• This pipeline adapts concepts and code from CT-TIQUA for MRI-based TBI analysis. Source is https://github.com/nifm-gin/CT-TIQUA
• Labled training data from https://dataverse.nl/dataset.xhtml?persistentId=doi:10.34894/E0U32Q with full reference: @data{E0U32Q_2024, author = {Hugo J. Kuijf and Edwin Bennink and Koen L. Vincken and Nick Weaver and Geert Jan Biessels and Max A. Viergever}, publisher = {DataverseNL}, title = {{MR Brain Segmentation Challenge 2018 Data}}, year = {2024}, version = {V1}, doi = {10.34894/E0U32Q}, url = {https://doi.org/10.34894/E0U32Q} }

License

This project is licensed under the MIT License - see the LICENSE file for details.

Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines on how to contribute to this project.