liaowc / nuclei-instance-segmentation

Nuclei Instance Segmentation

image source

Install

1. Install PyTorch

   pip3 install torch # if you haven't installed it

2. Install detectron2

   # See your PyTorch version
   python3 -c "import torch; print(torch.__version__)"
   # See your Cuda version
   nvcc --version

   Refer to: https://detectron2.readthedocs.io/en/latest/tutorials/install.html Find the instruction for installing dectectron2 based on your Cuda and PyTorch version.

   # E.g. if the Cuda version and the PyTorch version of current device is 11.1 and 1.10.0 respectively.
   # then you can:
   python3 -m pip install detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu111/torch1.10/index.html

3. Install other packages

   pip3 install torchvision
   
   pip3 install opencv-python

Prepare (Before Train or Inference)

1. Download the dataset.zip (72 MB), unzip it, and move it in the repo root path. The structure will be:

   Nuclei-Instance-Segmentation/
   |___dataset/
       |___test/
       |   |___TCGA-50-5931-01Z-00-DX1.png
       |   |___TCGA-A7-A13E-01Z-00-DX1.png
       |       ...... (other 4 images, 6 test images in total)
       |___train/
       |   |___TCGA-18-5592-01Z-00-DX1 (first data's directory)
       |   |   |___images
       |   |   |   |___TCGA-18-5592-01Z-00-DX1.png (this data's image)
       |   |   |___masks
       |   |       |___mask_0001.png
       |   |       |___mask_0002.png
       |   |           ...... (all masks of instances)
       |   |___TCGA-21-5784-01Z-00-DX1 (second data's directory)
       |        ...... (other data, 24 in total)
       |___test_img_ids.json
   ...... (other dir)
   

2. Run prepare.py to automatically produce train/valid/test dataset.
   Default: 24 train, 0 valid. If you want to change the number of valid images, modify the var in prepare.py.

   # Default setting
   N_VALID = 0 # 0 valid
   MAX_IMG_ID = 23 # 24 data in total, 0 for valid, 24 for train, so the max img id is 23 (0~23)

   # If you want 2 valid
   N_VALID = 2 # 0 valid
   MAX_IMG_ID = 21 # 24 data in total, 2 for valid, 22 for train, so the max img id is 21 (0~21)

   python3 prepare.py

3. Download my pretrained model: link
   This was trained 37999 iteration from the pretrained model of cascade_mask_rcnn_R_50_FPN_3x on detectron2.

   pip3 install gdown
   gdown --id 1to1K6_dnSVnY9-Qv3CJz_ITTaPqaxA1m

4. Make augmentation data: follow the instruction of make_augmentation_data.ipynb and run all the cells. Augmentation data will be produced and saved in "aug/". (If you want to inference only, this part can be skipped.)

Train

Train:

# 1. Modify the settings in main.py (e.g. from which pretrained model)
# 2. Run it.
python3 main.py

Track your training using Tensorboard:

# Default dir of training is "output/" 
tensorboard --logdir output

Inference

Inference test images and make a submit file.

# You can modify the setting in inference.py
python3 inference.py

Reference

• https://gilberttanner.com/blog/detectron2-train-a-instance-segmentation-model
• https://detectron2.readthedocs.io/en/latest/index.html
• https://github.com/072jiajia/CVDL_HW3
• https://github.com/joheras/CLoDSA
• https://www.immersivelimit.com/tutorials/create-coco-annotations-from-scratch