CRLM-instance

Code for ISBI-2025 Conference - INSTANCE-AWARE DEEP LEARNING FOR LIVER METASTASIS SEGMENTATION

Overview

This project focuses on investigating the use of instance-aware loss approach in deep segmentation models applied to the detection and segmentation of liver colorectal metastases (CRLM).

Key Features

• Automated liver and tumor segmentation using pre-trained SwinUNETR models
• Support for two model variants: standard SwinUNETR and instance-aware SwinUNETR
• Post-processing with volume-based filtering and connected component analysis
• Output segmentation masks with 3 classes (0: background, 1: liver, 2: tumors)

Input Requirements

• CT scans must be provided in NIfTI format (.nii.gz)
• Images will be processed in abdominal CT window settings
• Our models were trained on scans resampled to 1.5×1.5×2.0 mm³ voxels

Pre-trained Models

Pre-trained weights for two SwinUNETR variants are available:

• SwinUNETRb: Baseline model trained with standard Dice+CE loss
• SwinUNETR•: Enhanced model incorporating instance-aware loss for improved small lesion detection

Access to model weights is provided through our cloud archive https://drive.google.com/drive/folders/1II_7cp-it7lI2vnUGDgD33SSSHpaROuV?usp=sharing.

Usage

Configure the input/output paths and model selection:

directory = ""     # Input directory containing .nii.gz files
output_dir = ""   # Output directory for segmentation results
weights_path = ""  # Path to pre-trained model weights

Install the dependencies

pip install -r requirements.txt

Run the inference :

python inference.py

The script will:

1. Process all .nii.gz files in the input directory
2. Generate segmentation masks with labeled regions:

• 0: Background
• 1: Liver tissue
• 2: Tumor regions

3. Apply post-processing to filter small components and retain only significant lesions
4. Save the results as NIfTI files in the specified output directory

Performance Notes

As detailed in our paper, the models demonstrate different strengths:

• SwinUNETRb achieves balanced performance across lesion sizes
• SwinUNETR• shows improved detection of small lesions (<1000mm³)
• Both variants achieve high recall but may generate more false positives compared to traditional approaches (eg, nn-Unet).

Post-processing helps filter out small artifact components with a fixed volume threshold to reduce false positives generation.

Reference

If you use this code in your research, please cite our paper:

@inproceedings{Abbas2025Instance,
	author = {Abbas, M. and Andrade-Miranda, G. and Jaouen, V. and Bourbonne, V. and Bert, J. and Visvikis, D. and Badic, B. and Conze, P.-H.},
	booktitle = {2025 {IEEE} 22nd {International} {Symposium} on {Biomedical} {Imaging} ({ISBI})},
	doi = {10.1109/isbi60581.2025.10980698},
	year = {2025},
	month = {4},
	pages = {1--5},
	organization = {IEEE},
	title = {Instance-{Aware} {Deep} {Learning} for {Liver} {Metastasis} {Segmentation}},
	url = {http://dx.doi.org/10.1109/ISBI60581.2025.10980698},
}

Contact

marwan.abbas@univ-brest.fr