Skip to content

Latest commit

 

History

History
210 lines (167 loc) · 8.57 KB

runningApplication.md

File metadata and controls

210 lines (167 loc) · 8.57 KB
Raw

Running the Application

Note the ${fets_root_dir} from Setup.

Table of Contents

Application Path

cd ${download_location}
${fets_root_dir}/bin/FeTS # launches application

Please add the following path to your LD_LIBRARY_PATH when using FeTS: ${fets_root_dir}/lib:

export LD_LIBRARY_PATH=${fets_root_dir}/lib:$LD_LIBRARY_PATH

Back To Top ↑

Inference

Inference using BraTS-winning algorithms

${fets_root_dir}/bin/FeTS_CLI_Segment -d /path/to/output/DataForFeTS \ # data directory after invoking ${fets_root_dir}/bin/PrepareDataset
  -a deepMedic,nnunet,deepscan \ # all pre-trained models currently available in FeTS see notes below for more details
  -lF STAPLE,ITKVoting,SIMPLE,MajorityVoting \ # todo: select the most appropriate after Ujjwal's analysis
  -g 1 \ # '0': cpu, '1': request gpu
  -t 0 # '0': inference mode, '1': training mode

The aforementioned command will perform the following steps:

  • Perform inference on the prepared dataset based on selected architectures and label fusion strategies
    • NOTES:
      • nnUNet should have been downloaded during the installation process. In case there was a failure, follow these steps:
        • Download weights from this URL or from the command line: 
          cd ${fets_root_dir}/data/fets
wget https://upenn.box.com/shared/static/f7zt19d08c545qt3tcaeg7b37z6qafum.zip -O nnunet.zip
unzip nnunet.zip
        • Unzip it in ${fets_root_dir}/data/fets. The directory structure should look like this: 
          ${fets_root_dir}
│
└───data 
│   │
│   └──fets
│   │   │
│   │   └───nnunet
│   │   │   │
│   │   │   └───${training_strategy}.1_bs5 # the specific training strategy
│   │   │   │   │
│   │   │   │   └───fold_${k} # different folds
│   │   │   │   │   │ ...
        • This will now be available as a model for inference using the FeTS_CLI_Segment applications under the -a parameter.
      • To run DeepScan, at least 120G of RAM is needed.
      • DeepMedic runs as a CPU-only task.
  • Leverage the GPU
  • Place inference results on a per-subject basis for quality-control: 
    DataForFeTS
│
└───Patient_001 # this is constructed from the ${PatientID} header of CSV
│   │ Patient_001_brain_t1.nii.gz
│   │ Patient_001_brain_t1ce.nii.gz
│   │ Patient_001_brain_t2.nii.gz
│   │ Patient_001_brain_flair.nii.gz
│   │
│   └──SegmentationsForQC
│   │   │ Patient_001_deepmedic_seg.nii.gz # individual architecture results
│   │   │ Patient_001_nnunet_seg.nii.gz
│   │   │ Patient_001_deepscan_seg.nii.gz
│   │   │ Patient_001_fused_staple_seg.nii.gz # label fusions using different methods
│   │   │ Patient_001_fused_simple_seg.nii.gz
│   │   │ Patient_001_fused_itkvoting_seg.nii.gz
│
└───Pat_JohnDoe
│   │ ...

Back To Top ↑

Inference using the FeTS Consensus Models

${fets_root_dir}/bin/FeTS_CLI_Inference -d /path/to/output/DataForFeTS \ # data directory after invoking ${fets_root_dir}/bin/PrepareDataset
  -o /path/to/output/InferenceResults \ # output directory
  -g 1 \ # '0': cpu, '1': request gpu

The aforementioned command will run inference using the FeTS Consensus models (both singlet and triplet) for the data directory specified by -d. The output will be placed in the directory specified by -o.

Back To Top ↑

Manual Corrections

  • Use the FeTS graphical interface (or your preferred GUI annotation tool such as ITK-SNAP or 3D-Slicer) to load each subject's images:

    LoadImages

  • And each segmentation (either individual architectures or the fusions) separately:

    LoadROI

  • Perform quality control and appropriate manual corrections for each tumor region using the annotation tools:

    Annotation

    • Ensure BraTS annotation protocol is followed for labels:
    Label Region Acronym
    1 Necrotic Core of Tumor NET
    2 Peritumoral Edematous & Infiltrated Tissue ED
    4 Enhancing/Active part of tumor ET

  • Save the final tumor segmentation as ${SubjectID}_final_seg.nii.gz under the subject's directory:

    DataForFeTS
│
└───Patient_001 # this is constructed from the ${PatientID} header of CSV
│   │ Patient_001_brain_t1.nii.gz
│   │ Patient_001_brain_t1ce.nii.gz
│   │ Patient_001_brain_t2.nii.gz
│   │ Patient_001_brain_flair.nii.gz
│   │ Patient_001_final_seg.nii.gz # NOTE: training will not work if this is absent!!!
│
└───Pat_JohnDoe
│   │ ...

Back To Top ↑

Sanity Check

Before starting final training, please run the following command to ensure the input dataset is as expected:

cd ${fets_root_dir}/bin
./OpenFederatedLearning/venv/bin/python ./SanityCheck.py \
  -inputDir /path/to/output/DataForFeTS \
  -outputFile /path/to/output/sanity_output.csv

Note: If you are running FeTS version 0.0.2 (you can check version using FeTS_CLI --version), please do the following to get the SanityChecker for your installation:

cd ${fets_root_dir}/bin
wget https://raw.githubusercontent.com/FETS-AI/Front-End/master/src/applications/SanityCheck.py

Phase-2 Intensity Check

During discussions with some of the collaborating sites, we note negative values randomly coming up in the pre-processed scans. To identify these cases, we have now put together an additional script to assess all pre-processed images for the negative values and provide relevant statistics. This can be run in the following manner:

cd ${fets_root_dir}/bin
wget https://raw.githubusercontent.com/FETS-AI/Front-End/master/src/applications/Phase2_IntensityCheck.py
./OpenFederatedLearning/venv/bin/python ./Phase2_IntensityCheck.py \
  -inputDir /path/to/output/DataForFeTS \
  -outputFile /path/to/output/intensity_check.csv

If /path/to/output/intensity_check.csv doesn't get generated, that means the dataset is completely fine, otherwise, please send the file to admin@fets.ai for debugging.

Back To Top ↑

Training

Ensure Sanity Checking Is Done

Proceed to training once sanity check (↑) is successfully finished.

Transfer Certificates

If you have a signed certificate from a previous installation, ensure they are copied before trying to train:

cd ${fets_root_dir}/bin/
cp -r ${fets_root_dir_old}/bin/OpenFederatedLearning/bin/federations/pki/client ./OpenFederatedLearning/bin/federations/pki

Start Training

${fets_root_dir}/bin/FeTS_CLI -d /path/to/output/DataForFeTS \ # input data, ensure "final_seg" is present for each subject
  -c ${collaborator_common_name} \ # common collaborator name created during setup
  -g 1 -t 1 # request gpu and enable training mode

The aforementioned command will perform the following steps:

  • Train a model with a 3DResUNet architecture (more architectures will be made available in future releases) for all the complete subjects (i.e., with ${SubjectID}_final_seg.nii.gz and all 4 structural modalities present) in a collaborative manner
  • Leverage the GPU
  • Perform full-image validation during training

Back To Top ↑

Next: Contact FeTS Admin