search

ivadomed / model-spinal-rootlets

Deep-learning based segmentation of the spinal nerve rootlets
5 stars 2 forks source link

How to use nnUNet (V2) a quick start guide #6

Closed tzebre closed 1 year ago

tzebre commented 1 year ago

What is nnUNet ?

nnUNet is a segmenting framework with automatic data-driven augmentation available on github.

Installation

Code management

need python>=3.9

mkdir nnUNet_env
python -m venv nnUNet_env/
source nnUNet_env/bin/activate
git clone https://github.com/MIC-DKFZ/nnUNet.git
cd nnUNet
pip install -e .

Run this on GPU cluster, so install this in ~/

Data management

You have to create 3 directory (copy paste this code)

mkdir {nnUNet_preprocessed, nnUNet_raw, nnUNet_results}
echo add this next 3 lines to your .bash profile 
echo 'export nnUNet_raw=$(pwd)"/nnUNet_raw"' 
echo 'export nnUNet_preprocessed=$(pwd)""/nnUNet_preprocessed"' 
echo 'nnUNet_results=$(pwd)"/nnUNet_results"'

if using rosenberg do it in nvme or extra_ssd

nnUNet_raw

Place the raw datasets This folder will have one subfolder for each dataset names DatasetXXX_YYY where XXX is a 3-digit identifier (such as 001, 002, 043, 999, ...) and YYY is the (unique) dataset name.

nnUNet_raw/Dataset001_NAME1
├── dataset.json
├── imagesTr
│   ├── BRATS_001_0000.nii.gz
│   ├── BRATS_001_0001.nii.gz
│   ├── ...
├── imagesTs
│   ├── BRATS_485_0000.nii.gz
│   ├── BRATS_485_0001.nii.gz
│   ├── ...
└── labelsTr
    ├── BRATS_001.nii.gz
    ├── BRATS_002.nii.gz
    ├── ...

dataset.json

Contains metadata that nnU-Net needs for training

{ 
 "channel_names": { # Remove unused channels
   "0": "FLAIR", 
   "1": "T1w",
   "2": "T2",
   "3": "T2w"
 }, 
 "labels": {  # Use label values from 0 to X
   "background": 0,
   "label_1": 1,
   "label_2": 2
 }, 
 "numTraining":  00, # replace by number of subject in imagesTr 
 "file_ending": ".nii.gz"
 "overwrite_image_reader_writer": "SimpleITKIO"  
 }

imagesTr

Contains the images belonging to the training cases. NAME_NUM_CHANNEL.nii.gz NUM = subject number CHANNEL = FLAIR (0000), T1w (0001), T1gd (0002) and T2w (0003)

labelsTr

Images with ground truth segmentation NAME_NUM.nii.gz NUM = subject number

imagesTs

Folder for test images, not used by nnUNet.

Train a model

Verify data integrity

nnUNetv2_plan_and_preprocess -d DATASET_ID --verify_dataset_integrity

Exemple for Dataset001_NAME1 replace DATASET_ID by 001

Train model

CUDA_VISIBLE_DEVICES=X nnUNetv2_train DATASET_ID CONFIG FOLD --npz

Replace X by GPU id for training Every 50 epoch a checkpoint is saved (dont stop before the 50th epoch if you wan to run inference)

Configuration

Track progress

Graph tracking progress available in nnUNet_results/DATASET_ID/nnUNetTrainer__nnUNetPlans__3d_fullres/fold_X/progress.png

You can copy locally from folder with scp PATH:server_file PATH:local_file

Run inference

Basic run

Only possible if 50+ epoch nnUNetv2_predict -i nnUNet_raw/DATASET_ID/imagesTs -o OUT_DIR -d DATASET_ID -c CONFIG --save_probabilities -chk checkpoint_best.pth -f FOLD

If you've train a 3d_fullres on fold_0 with Dataset_001 use nnUNetv2_predict -i nnUNet_raw/Dataset001_rootseg/imagesTs -o result_test/ -d 001 -c 3d_fullres --save_probabilities -chk checkpoint_best.pth -f 0

Best configuration run

...

tzebre commented 1 year ago

If the train did not go to the end (1000 epochs), the best configuration method is not usable. About epoch num change: https://github.com/MIC-DKFZ/nnUNet/issues/1384#issuecomment-1531177506

tzebre commented 1 year ago

Documentation available here