Skip to content

Latest commit

 

History

History
162 lines (111 loc) · 10.1 KB

README.md

File metadata and controls

162 lines (111 loc) · 10.1 KB

Deep Learning and Multi-Modal MRI for Segmentation of Sub-Acute and Chronic Stroke Lesions

This project is for research and educational purposes only. Commercial use of the models or data is strictly prohibited unless prior permission is granted by the authors.

Research article under revision
Conference paper: Accepted and presented at the 13th World Congress for Neurorehabilitation (WCNR 2024).

Contact: Francesca Galassi (francesca.galassi@irisa.fr)

Overview

This repository contains:

  1. Trained Models:

    • A baseline single-modality model trained on the ATLAS v2.0 dataset (T1-w).
    • A fine-tuned single-modality (T1-w) model trained on a private dataset.
    • A fine-tuned dual-modality (T1-w + FLAIR) model trained on a private dataset.

  2. Code: End-to-end pipeline for segmentating your data in test mode.

  3. Additional Scripts:

    • Fine-tuning script: Adapts a single-modality nnU-Net model to handle two modalities (T1-w and FLAIR).

Methods and Materials

Datasets

Data used in this project include a publicly available single-modality dataset (ATLAS v2.0) and an internal dual-modality dataset, combining T1-w and FLAIR MRI scans.

1. Public Single-Modality ATLAS v2.0 Dataset

The ATLAS v2.0 dataset was used to develop the baseline single-modality model. It includes T1-w MRI scans and corresponding lesion segmentation masks, pre-aligned to the MNI-152 standard template with a voxel size of 1 x 1 x 1 mm. The dataset used for training and validating our model consists of 655 T1-w MRI scans from the training folder of the ATLAS v2.0 dataset. The ATLAS v2.0 dataset is publicly available for download here.

2. Internal Dual-Modality Datasets

The internal dual-modality datasets were collected from two clinical studies: the NeuroFB-AVC study (NCT03766113) and the AVCPOSTIM study (NCT01677091). These datasets, which include both T1-w and FLAIR MRI scans, were used for fine-tuning the baseline model with dual-modality inputs, as well as for testing. Manual segmentation of the FLAIR images was carried out by a neuroimaging expert, with assistance from the T1-w modality, and reviewed by a neuroradiologist. For access to the internal datasets, please contact the authors and refer to the corresponding clinical study registrations:

Ethical Considerations: all subjects in the internal datasets provided written informed consent before participation. The studies were approved by the relevant ethics committees and complied with French data confidentiality regulations.

Preprocessing Pipeline

The preprocessing pipeline is a critical component in the preparation of MRI data prior to segmentation model training and testing. These preprocessing steps were applied to all MRI data involved in the project, including both the internal datasets and the ATLAS dataset. When running the model in test mode, you should perform the preprocessing steps (1-5) on your data first.

1. Brain Extraction

The HD-BET tool is used to remove the skull from the images. This deep learning-based method provides improved accuracy for brain extraction compared to traditional methods. For more information on HD-BET, please refer to https://github.com/MIC-DKFZ/HD-BET.

2. Re-orientation

The volumes are re-oriented to the RAS (Right-Anterior-Superior) coordinates to ensure consistent orientation across all images. This step is crucial for standardizing image orientation and preventing issues during further processing.

3. Registration

If both T1-w and FLAIR images are available for a subject, the T1-w image is rigidly registered to the corresponding FLAIR image using a block matching registration method (animaPyramidalBMRegistration, Anima (https://anima.irisa.fr/)). If only the T1-w modality is available, this step is skipped. This ensures alignment between T1-w and FLAIR images when both are present, supporting multi-modality analysis.

4. Bias Correction

The bias due to spatial inhomogeneity is estimated and removed from the data using the N4ITK bias field correction algorithm (animaN4BiasCorrection). This correction compensates for signal variations caused by magnetic field inhomogeneities during MRI acquisition, ensuring more uniform intensity across the image.

5. Intensity Normalization

Image intensities are standardized by subtracting the mean voxel value and dividing by the standard deviation for each image. This normalization step ensures that all images have a consistent intensity distribution, making them comparable across subjects and modalities, which is crucial for downstream analyses and model training.

Note on nnU-Net Framework

It is important to note that in addition to the preprocessing steps outlined above, the nnU-Net framework (https://github.com/MIC-DKFZ/nnUNet/tree/nnunetv1) performs several additional operations by default to standardize the input data:

  • Voxel Spacing Standardization: The framework automatically calculates a uniform target spacing for each axis using the median values from the training dataset. This ensures that all images are resampled to have consistent voxel spacing across subjects and modalities.

  • Resampling Strategy: By default, third-order spline interpolation is used for resampling images. However, for anisotropic images (where the ratio of the maximum to minimum axis spacing exceeds 3), the framework adapts its resampling approach:

    • In-plane resampling is done using third-order spline interpolation.
    • Out-of-plane resampling uses nearest neighbor interpolation to reduce resampling artifacts and better preserve spatial information.

  • Segmentation Maps: The framework automatically converts segmentation maps into one-hot encodings. Each channel of the segmentation mask is then interpolated using linear interpolation, and the final segmentation mask is obtained by applying the argmax operation. For anisotropic data, nearest neighbor interpolation is applied to the low-resolution axis to minimize interpolation artifacts.

This preprocessing ensures consistency across all images and improves the robustness of the model, especially when dealing with different image resolutions and anisotropic data.

Test Mode Example

Before running the commands, ensure your data is organized and structured correctly as follows:

<input_data_folder>/
│
└── raw/
    │
    └── patient01/
        ├── anatomy-brain
  • The raw folder should contain a folder for each patient (e.g., patient01), with subfolder anatomy-brain.
    • The anatomy-brain subfolder should contain the modalities for each patient (e.g., flair_skull_stripped.nii.gz, t1_skull_stripped.nii.gz).

An example is provided within the test_mode folder - simple and straightforward.

Note: you need to perform step 1 of the preprocessing pipeline before running the following commands. Indeed, the first command performs preprocessing steps (2-5).

To run the code in test mode, you can use the following command examples. These commands are designed to predict on a batch of data.

python3 preprocess.py --patients <input_data_folder>/raw/ --preprocess_steps prepare_without_brain_extraction remove_bias normalize -cs -m t1 -cfg config_vtest.yml 

python3 predict_batch.py --patients <input_data_folder>/normalized/ --output <input_data_folder> -cs -mn Task111 -m t1 --preprocess_steps check install -pmin <min_value> -pmax <max_value> -vmin <validation_min> -cfg config_vtest.yml -f 2

In our study we used the following parameters: -pmin 0.2 -pmax 0.2 -vmin 10

The evaluation script uses animaSegPerfAnalyzer for lesion detection performance analysis (https://anima.readthedocs.io/en/latest/segmentation.html).

Available models :

  • Task111 :The baseline model trained on the ATLAS v2.0 dataset (T1-w). Folder: 2.
  • Task999 :The fine-tuned model on T1-w. Folder: 4.
  • Task777 :The fine-tuned model on both T1-w and FLAIR. Folder: 2.

Installation and Requirements

This section provides a step-by-step guide on how to install and run MM-StrokeNet. Before proceeding, ensure your system meets the following requirements:

Software Requirements

  • Python 3.9 or higher
  • Torch 2.0.0 or higher
  • Scipy
  • NumPy
  • scikit-learn
  • scikit-image 0.19.3 or higher

Hardware Used

  • NVIDIA GPU (RTX A5000 GPU and 30,6GiB RAM) with CUDA 11.4

Installation Steps

Clone this repository with the following command:

git clone https://github.com/LounesMD/MM_StrokeNet.git


## MM-StrokeNet packages
In this repository you will find the following folders :
* Algorithm : Custom scripts used to modify the pretrained nnU-Net model for compatibility with dual-modality (T1-w + FLAIR) inputs.
* Images : Visual assets used in the paper, including images and figures.
* Models : Pretrained and fine-tuned model weights and configurations. This includes:
	*   The baseline model trained on the [ATLAS v2.0 dataset](https://fcon_1000.projects.nitrc.org/indi/retro/atlas.html).
	*   The fine-tuned model on T1-weighted MRIs.
	*   The fine-tuned model on both T1-weighted and FLAIR MRIs.

## Paper :
The original research paper is currently under review.
Initial results were presented in the form of an oral presentation at the 13th World Congress for Neurorehabilitation (WCNR) 2024. Here is a [link](https://hal.science/hal-04546362) to the paper's abstract.

## Citing 
If you use this project in your work, please consider citing it as follows:

```bibtex
@misc{MM-STROKEnet,
  authors = {Lounès Meddahi, Stéphanie s Leplaideur, Arthur Masson, Isabelle Bonan, Elise Bannier Francesca Galassi},
  title = {Enhancing stroke lesion detection and segmentation through nnU-net and multi-modal MRI Analysis},
  year = {2024},
  conference = {WCNR 2024 - 13th World Congress for Neurorehabilitation, World federation for Neurorehabilitation},
}