bladder_segmentation

Guided Research for Bladder Segmentation from CT and PET images

Main points:

• use resize_amos.py, resize_ctorg.py, then merge_amos_ctorg.py to create the merged dataset
• you can play around with the spacings/dimensions in those files
• the output folder should be directly runnable with nnUNet

Prerequisites:

• a training setup (Polyaxon / what else is currently being used)
• having write access to your NAS directory
• having access to the raw AMOS and CT-ORG datasets in the NAS
• wandb

Explanation for each of the files:

wandb_login.py

I created a file name wandb_login.py and called it to login into wandb at every run, so I don't have to post my private credentials anywhere It's just a python script you call before running the training, and it has two lines: import os os.system("wandb login --relogin <your_token>") With the logging I set up, your dashboard should look like this

resize_amos and resize_ctorg

Both do the preprocessing steps on the 3d volumes and their respective labels. They take the (volume,label) pair from their initial folders, create a segmentation mask (the initial volumes for both CT-ORG and AMOS have multiple classes)

train.py

An adaptation of https://github.com/Project-MONAI/tutorials/blob/main/3d_segmentation/spleen_segmentation_3d_lightning.ipynb

3d_vis.ipynb

Can use to visualise slices if you don't have ImFusion

Extras

• in the polyaxonfile.yaml I specified some GPUs are upset at the pytorch version being used
• you can search for pytorch images here https://hub.docker.com/r/pytorch/pytorch/tags. MONAI (https://hub.docker.com/r/projectmonai/monai/tags) took too long for me to download
• i recommend not putting polyaxon-cli in your requirements.txt. The version used at IFL is older and it has some dependency conflicts with other packages such as wandb/jupyter
• pred.py is patched together so it can run on the folder format i had on the internal data. Modify it as needed