2D Vesicle Instance Segmentation

Main framework of training and inference code comes from C-3-Framework. The original liscense is included under ./liscense

Environment

• Follow C-3-Framework
  • Python 3.x
  • Pytorch 1.0 (some networks only support 0.4): http://pytorch.org .
  • other libs in requirements.txt, run pip install -r requirements.txt.
  • Perhaps some packages like h5py need to be installed after above command. pip install h5py

Data Preparation

• cd ./data
• Modify settings in ./data/config.py
• Run python 0_augment.py to do augmentation manually.
• Run python 1_data_prepared.py to split sugmented data into train, valid, test datasets.

Training

• Modify settings in ./training/config.py
• cd ./training
• python train.py
• Results will be stored at ./training/exp/. Tensorboard can be used to visualize the result by tensorboard --logdir=exp --port=6006

Pretrained Model

• Pretrained model can be obtained from google drive
• Download to default position at ./model

Inference

cd ./inference

1. Synapse locations to input images and masks.
   • By python 0_synapse_gen.py
   • Output to OUTPUT_DIR defined in ./code/config.py, including:
     • dir full: full input image
     • dir part: masked input image (not used)
     • dir mask: binary mask
2. Use trained model to predict the heat map.
   • By python 1_test_all.py.
   • Predict with MODEL_DIR/MODEL_NAME directed by ./code/config.py. Output to OUTPUT_DIR/result, including:
     • dir pred: heat map in npy format
     • dir mask: predicted mask by 0.3 threshold
     • dir gt: empty
3. Multiply the predicted mask and heat map with input mask.
   • By python 2_multiply.py
   • Output to OUTPUT_DIR/mask_result, including:
     • dir pred: heat map in npy format
     • dir mask: predicted binary mask by 0.3 threshold.
4. Use watershed to cut heat map into vesicles. (Two-step Watershed)
   • By python 3_post_process.py
   • Output to OUTPUT_DIR/wshed_result, including:
     • dir data: segmentations in npy format
     • dir img: matplotlib images of segmentation result
5. Analyze the final 3D bounding box of synapse seg and prepare to display in VAST.
   • By python 4_final_step.py
   • Output to OUTPUT_DIR/vast_volume