edoardogiacomello / segan-cat

SegAN-CAT

Environment setup instructions

The project is provided as a repository of python scripts and a pre-configured docker environment running Jupyter Lab.

For running the environment with GPU support you will need nvidia-docker, cuda=10.0

1 - Clone the repo 2 - Within the repo root folder launch the script build_docker_[cpu|gpu].sh according to your gpu availability. This will build a docker image tagged edoardogiacomello/deepmri2:latest-[gpu|cpu] 3 - For running the environment for the first time, launch the script first_run_[cpu|gpu].sh according to your gpu availability. This will run a container named deepmri2. 4 - run tmux for managing multiple bash sessions. 5 - Run Jupyter lab with jupyter lab --ip=0.0.0.0 --allow-root

Jupyter will be available at :8887. You can optionally run tensorboard on port 6006.

Use ctrl+p+q to detach from the docker environment. Use docker attach deepmri2 to re-attach to the environment.

Paths description

• train_<arch>_<dataset>_options.py: Scripts for training the base models
• transfer_<arch>_<dataset>_options.py: Scripts for performing transfer learning. Edit your script to your needs
• dataset_helpers.py: Helper functions for managing the dataset (both conversion/pre-processing from raw to .tfrecords and to load the dataset during training/evaluation)
• DeepMRI.py main class that implements the training/validation/testing process.
• SegAN_IO_arch.pydefinition of our proposed architecture to be used in DeepMRI.py
• SegAN_no_dice_arch.pydefinition of the baseline architecture to be used in DeepMRI.py
• SegAN_arch.pydefinition of the SegAN \w Dice Loss architecture to be used in DeepMRI.py
• TestScores.ipynb Notebook for evaluating the networks.
• DatasetPreprocessing.ipynb: Example of how to run the pre-processing to convert from .mha to .tfrecords
• MajorityVoting*.ipynb Notebook for performing majority voting between multiple models
• NegotiationTools.py class needed for Majority voting experiments
• PredictOnChallenge.ipynb notebook for end-to-end prediction from .mha files. Saves the results in the brats2015 name format
• /models/<run_name>/: This folders will be created upon the first training. It contains the checkpoints, a csv for training and validation steps and the tensorboard data for each run.

segan-cat's Issues

Add HGG/LGG Classification

Import Brats2019 Dataset

• Review the updates on the dataset
• Check the follow-ups, consider splitting on a patient base, possibly stratifying also on real patients vs synthetic

Add multilabel segmentation

This will require to redefine the evaluation metrics:

• TF evaluation will be done per-slice and per-label (output a vector [batch, merics, labels] )
• They will be merged afterward (on a MRI basis) according to the BRATS evaluation metrics

Check if it's feasible to adopt WGAN-GP

Sumbit the results to Brats2015

• Check if cropping has an impact on performances
• Sumbit on Training for preliminary results

Port correct Evaluation also to Model Selection

Per-MRI Evaluation should also be used for model selection during training.

• Fix or Replace Logger class with an evaluation procedure similar to the current evaluate() behavior (based on DataFrames for aggregating metrics over each MRI).
• Review save and load functions.
• Eventually move to TensorBoard for easier inspection (now it can be run inside the notebook itself)
• Simplify unnecessary Progressbars and metrics running averages (It's not necessary to show averages on progressbars, just current batchwise metrics for easier debug)
• Update evaluation() function