MBH-Seg25: Multi-class Brain Hemorrhage Segmentation in Non-contrast CT

Description

Inference repository for the 5th solution in the second edition of the MICCAI MBH-Seg25 Challenge on multi-class brain hemorrhage segmentation from non-contrast CT scans.

challenge website: https://www.mbhseg.com/

leaderboard: leaderboard

Model Details

File Introduction

• MBH-Seg25: contains the predict code, model, and weights for MBH-Seg25.
• mbh_seg.tar: Docker image for MBH-Seg25.

Quick Start

1. Install Dependencies

# Load the Docker image
docker load -i mbh_seg.tar 

# Run the Docker container
docker run -itd --gpus all --name mbh_seg mbh_seg:latest 

# Copy your data、code、weights to the container
docker cp /path/to/MBH-Seg25 mbh_seg:/workspace

# Access the container
docker exec -it mbh_seg -w /workspace /bin/bash 

# Install the dependencies
cd /workspace/MBH-Seg25/selective_scan
pip install -e .

2. Prepare Test Data

MBH_Test_label_2025/
    ├── ID_90ae3af3_ID_8d77fcb5d2/
    │   ├── image.nii.gz
    │   ├── label_annot_1.nii.gz
    │   └── label_annot_5.nii.gz
    ├── ID_066b1fc2_ID_f937d7bff0/
    │   ├── image.nii.gz
    │   ├── label_annot_1.nii.gz
    │   └── label_annot_5.nii.gz
    └── ID_0219ef88_ID_e5c1a31210/
        ├── image.nii.gz
        ├── label_annot_1.nii.gz
        └── label_annot_5.nii.gz

• folder name: ID of the patient
• image.nii.gz: The input image for segmentation.
• label_annot_*.nii.gz: The ground truth label for the annotation.

# Copy test data to the container
docker cp /path/to/MBH_Test_label_2025 mbh_seg:/workspace/MBH-Seg25

3. Predict Test Data

# Access the container
docker exec -it mbh_seg -w /workspace /bin/bash 

cd /workspace/MBH-Seg25
python stroke_predict.py --input_dir /workspace/MBH-Seg25/MBH_Test_label_2025 --output_dir /workspace/MBH-Seg25/predictions

4. Evaluation with Evaluation Tool

The toolkit automatically detects your data structure and parameters.

ID-based structure (like MBH dataset)

project/
├── predictions/
│   ├── ID_90ae3af3_ID_8d77fcb5d2.nii.gz
│   ├── ID_066b1fc2_ID_f937d7bff0.nii.gz
│   └── ID_0219ef88_ID_e5c1a31210.nii.gz
└── MBH_Test_label_2025/
    ├── ID_90ae3af3_ID_8d77fcb5d2/
    │   ├── image.nii.gz
    │   ├── label_annot_1.nii.gz
    │   └── label_annot_5.nii.gz
    ├── ID_066b1fc2_ID_f937d7bff0/
    │   ├── image.nii.gz
    │   ├── label_annot_1.nii.gz
    │   └── label_annot_5.nii.gz
    └── ID_0219ef88_ID_e5c1a31210/
        ├── image.nii.gz
        ├── label_annot_1.nii.gz
        └── label_annot_5.nii.gz

Note: Prediction filenames must match the folder names in the labels directory for correct evaluation.

For most evaluation needs, simply use:

# Access the container
docker exec -it mbh_seg -w /workspace /bin/bash 

cd /workspace/MBH-Seg25
python competition_evaluation.py --pred_path /workspace/MBH-Seg25/predictions --gt_path /workspace/MBH-Seg25/MBH_Test_label_2025

This command automatically:

• Detects all prediction files in your predictions directory
• Matches them with corresponding annotation folders
• Evaluates using appropriate metrics
• Handles ID-based naming (like MBH dataset structure)

Evaluation Metrics

Diverse Performance Metrics (Set-Level)

• GED (Generalized Energy Distance): Set-level similarity between prediction and annotation sets
• Dice_max: Maximum Dice score achievable through optimal matching
• Dice_match: Dice score using Hungarian algorithm optimal matching
• Dice_soft: Soft Dice coefficient accounting for all annotation variations

Personalized Performance Metrics

• Individual Dice Scores: Per-annotator Dice coefficients
• Statistical Summary: Mean, standard deviation, min/max across annotators

Data Format Requirements

• File Format: NIfTI (.nii.gz)
• Segmentation Maps: Integer class indices (0, 1, 2, ...) or binary (0, 1)
• Spatial Consistency: All files must have matching dimensions
• Class Handling: Supports sparse class indices (e.g., [0, 2, 5])
• ID Matching: For MBH-style datasets, prediction filenames must exactly match the corresponding annotation folder names

Output

Final evaluation results for all samples are saved in the competition_results/ folder:

• competition_aggregate_results.json - Summary metrics across all samples
• competition_detailed_results.json - Detailed per-sample results
• competition_report.txt - Human-readable evaluation report