Medical Image Segmentation

Medical image segmentation using deep learning methods

Library

Please recovery environment by conda

conda env create -f pyenv.yml

or by pip

pip install -r requirements.txt

Usage

Change working directory to project root directory and activate virtual environment.

cd ~/MedicalImageSegmentation
source activate <env-name>

1. Process LiTS dataset

• Convert 3D nii images and labels to 2d slices:

PYTHONPATH=./ python DataLoader/Liver/extract.py

# -------------------------------------------------------
Please choice function:
	a: exit()
	b: run_nii_3d_to_png()
	c: run_dump_hist_feature()
	d: run_simulate_user_prior [A/b/c/d] b
# Choice `b` for executing function `run_nii_3d_to_png()`

2. Train/Evaluate a UNet model

A. For help

python ./entry/main.py --help

B. Write your bash file in ./run_scripts/my_config.sh.

You can copy and modify provided template bash file in ./run_scripts/template/001_unet.sh.

• Required Parameters:

  • --mode: {train, eval}
  • --tag: (better to use current script name, such as ${BASE_NAME%".sh"})
  • --model: {UNet}
  • --classes: (your class names, such as "Liver Tumor")
  • --test_fold: which fold for validating, others for training

• Optional Parameters:

  • --random_flip: 3 --noise_scale 0.05 (data augmentation)
  • --im_height: 256 --im_width 256 (specify training image size, -1 if not specified)
  • --im_channel: 3 (specify image channel)
  • --num_of_total_steps: (number of max training steps)
  • --primary_metric: ("/", such as "Liver/Dice")
  • --weight_decay_rate: (weight decay rate)
  • --learning_policy: (learning rate decay policy, such as "plateau")
  • --warm_start_from: "004_triplet/best_model.ckpt-215001" (absolute or relative path)

• Other Parameters:

  • Please run python ./entry/main.py --help for details

C. Add execution permission for your bash file

chmod u+x ./run_scripts/my_config.sh

D. Begin train/evaluate model

• training

./run_scripts/my_config.sh train 0
./run_scripts/my_config.sh train 0,1   # Using multiple GPUs

• evaluation

# evaluation with best checkpoint
./run_scripts/my_config.sh eval 0 --load_status_file checkpoint_best
# evaluation with final checkpoint
./run_scripts/my_config.sh eval 0 --eval_final
# evaluating and saving predictions
./run_scripts/my_config.sh eval 0 --save_predict

• other parameters...

3. Train/Evaluate GUNet model

A. For help

python ./entry/main_g.py --help

B. Write your bash file in ./run_scripts/my_config_v2.sh.

You can copy and modify provided template bash file in ./run_scripts/template/001_gunet.sh.

• Required Parameters: The same as UNet.

• Optional Parameters: The same as UNet.

• Some Extra Parameters:

  • --use_context: Use context guide
  • --context_list: Paired context information: (feature name, feature length). For example: hist, 200
  • --hist_noise: Add noise to histogram context
  • --hist_noise_scale: Histogram noise random scale
  • --use_spatial: Use spatial guide
  • --spatial_random: Probability of adding spatial guide to current slice with tumors when use_spatial is on
  • --spatial_inner_random: Random choice tumors to give spatial guide inside a slice with tumors
  • --save_sp_guide: Save spatial guide when evaluating

Then add execution permission for your bash file & begin train/evaluate model the same as UNet.

4. Deployment

• Use CPU

docker run -t --rm -p 8501:8501 -v "D:/Library/Documents/0WorkSpace/MedicalImageSegmentation/export_path/112_nf_sp_dp:/models/112_nf_sp_dp" -e MODEL_NAME=112_nf_sp_dp tensorflow/serving &

• Use GPU (require docker nvidia/cuda)

docker run --runtime=nvidia -t --rm -p 8500:8500 -p 8501:8501 -v "/media/Data/Library/Documents/0WorkSpace/MedicalImageSegmentation/export_path/115_nf_both1_v2:/models/115_nf_both1_v2" -e MODEL_NAME=115_nf_both1_v2 tensorflow/serving:latest-gpu &

• Use GPU (with custom driver installation)

docker run --runtime=nvidia -t --rm  -p 8500:8500 -p 8501:8501 -v "/usr/lib/nvidia-396:/custom/nvidia-396" -v "/home/mgh3dqi/zjw/MIS/export_path/112_nf_sp_fix:/models/112_nf_sp_fix" -e LD_LIBRARY_PATH=/custom/nvidia-396 -e NVIDIA_VISIBLE_DEVICES=1 -e MODEL_NAME=112_nf_sp_fix tensorflow/serving:latest-gpu-cuda9.0

Need implemented

Base

• Data loader
• Training routine
• Summaries
• Metrics
• Loss functions
• Evaluate
• Predict
• Visualization
• Multi-GPU Training
• Multi-GPU Evaluation/Prediction

Networks

• UNet

• Attention UNet

• GUNet

• G-AttentionUNet

• DenseUNet

• G-DenseUNet

• Deep Supervised UNet

• Pre-trained backbone

• Context guide

• Histogram context

• Other context

• Spatial guide