Brain Metastasis Detection

Deep learning-based Metastasis Detection to Enhance Reproducibility and Reduce Workload in Brain Metastasis Screening with MRI: A Multi-center Simulation Study

Summary statement

• Deep learning-based detection of brain metastasis has the potential to result in a more reproducible and standardized assessment and reduce reading time.

Key results

• Our deep learning-based system exhibited a robust detection sensitivity of 90.2% and positive predictive value of 88.2%, with multi-center validation.
• Agreement among readers regarding the number of metastases detected was greater with the deep-learning system than without it (limits of agreement, −0.281 vs. −0.163).
• The reading time was significantly reduced from mean 66.9 s to 57.3 s with the deep-learning system, regardless of the imaging center.

Table of Contents

• Brain Metastasis Detection
• Table of Contents
• MRI Acquisition Protocol
• Network Architecture
• Installation
• Usage
• Acknowledgements

MRI Acquisition Protocol

Center 1

The imaging parameters for the GRE-T1WI sequence were as follows: repetition time (TR)/echo time (TE), 9.8/4.6 ms; flip angle, 8°; field of view, 24 cm; section thickness, 1 mm; matrix, 1024 × 1024; and acquisition time, 2 minutes 58 s. The parameters for the fast spin-echo sequence were as follows: TR/TE, 600/28.4 ms; flip angle, 90°; field of view, 24 cm; section thickness, 1 mm; matrix, 512 × 512; and acquisition time, 3 minutes 33 s. The improved motion-sensitized driven-equilibrium pre-pulse consisted of one 90° excitation pulse, two 180° refocusing pulses, and one 90° excitation pulse with motion-sensitized gradients between radiofrequency pulses. The duration between the two 90° pulses (TEprep) was 28.3 ms, and the flow velocity encoding for gradient pulses was 3 cm/s.

Center 2

The parameters for the fast spin-echo sequence were as follows: TR/TE, 600/28.4 ms; flip angle, 90°; field of view, 24 cm; section thickness, 1 mm; matrix, 240 × 240; and acquisition time, 3 minutes 23 s. The improved motion-sensitized driven-equilibrium pre-pulse consisted of one 90° excitation pulse, two 180° refocusing pulses, and one 90° excitation pulse with motion-sensitized gradients between radiofrequency pulses. The duration between the two 90° pulses (TEprep) was 28.3 ms, and the flow velocity encoding for gradient pulses was 1.3 cm/s. The imaging parameters for the GRE-T1WI sequence were as follows: repetition time (TR)/echo time (TE), 5.9-8.6/2.8-4.7 ms; flip angle, 8°; field of view, 24 cm; section thickness, 1 mm; matrix, 240 × 240; and acquisition time, 3 minutes 2 s.

Table

Imaging parameters of 3D CET1WI sequence of two centers

Network Architecture

Network architecture of the deep learning system for detection and count of brain metastasis. The trained U-Net architecture and more detailed information such as feature map size, kernel, and strides. Since the input image size varies for each case, several patches are generated using a sliding window approach. Model prediction results for each patch overlap by half of the size of a patch and are aggregated to generate the final lesion mask.

For more information about nnU-Net, please read the following paper:

Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J., & Maier-Hein, K. H. (2020). nnU-Net: a self-configuring method 
for deep learning-based biomedical image segmentation. Nature Methods, 1-9.

Installation

Installed requirements can cause potential conflicts with your library. Therefore, we recommend that you configure the virtual environment to run the following process.

This program has been confirmed to work in Ubuntu 18.04 and 20.04. CUDA and cuDNN are required prior to installation.

1. This program was written using Python 3. Check your Python version.

2. Install PyTorch as described on their website (conda/pip).

3. Verify that a recent version of pytorch was installed by running

   python -c 'import torch;print(torch.backends.cudnn.version())'
   python -c 'import torch;print(torch.__version__)'   

   This should print 8200 and 1.11.0+cu113 (Apr 1st 2022)

4. Install nnU-Net

   git clone https://github.com/jieunp/BM_detection_AI.git
   cd BM_detection_AI
   pip install -r requirements.txt
   pip install -e .

5. Your 3D brain MR image data should be prepared in the following structure. The prefix means a unique idnetifier, such as the number of the sample, and the suffix after the underbar means a black blood image for 0000 and a T1 contrast enhanced image for 0001. All data must be in nifti (*.nii) format. A segmentation mask named only with a prefix is automatically extracted to the folder specified as the output folder.

    BM_detection_AI/
    ├── dataset
    │   ├── Patient_A_0000.nii   <- Black Blood image
    │   ├── Patient_A_0001.nii   <- T1 Contrast Enhanced image
    │   ├── Patient_B_0000.nii
    │   └── Patient_B_0001.nii
    └── output
        ├── Patient_A.nii        <- Segmentation mask
        └── Patient_B.nii
   

6. Download the model best.model file through the following download, and move it to BM_detection_AI/results/nnUNet/3d_fullres/meta/nnUNetTrainer__nnUNetPlans/all.

Usage

Detect metastasis by running predict_simple.py. -i is the input dataset folder name and -o is the folder name to export segmentation mask.

python nnunet/inference/predict_simple.py -i dataset -o output -t meta -tr nnUNetTrainer -m 3d_fullres -f all

Acknowledgements

This code borrows from

• MIC-DKFZ/nnUNet