This repository contains all the code that was needed to run mask rcnn algorithm over the training data of brain/tissue slices as precised in the report.
Please be sure to have all the following packages installed and working:
- tensorflow
- keras
- skimage
If you have the tensorflow-gpu implementation, be sure to have everything ready following this guide: https://towardsdatascience.com/setup-an-environment-for-machine-learning-and-deep-learning-with-anaconda-in-windows-5d7134a3db10 To perform the training, you would need a powerful enough gpu. We used paperspace, setting it up following this guide: https://github.com/BrainSegmentation/lipschitz-lizards/wiki/Paperspace-Setup which was kindly provided by another group working with us.
For the visualization part, we have made a slight modification to the visualization function in Mask_rcnn. Clone Mask RCNN from here: https://github.com/matterport/Mask_RCNN/ and copy our visualize.py in their mrcnn folder. This modification is needed for single patch validation but not for the whole image results. It's used to have 2 channels visualization which is normally not supported. After modifying the file, you can install mask rcnn in your machine (python setup.py install in mask rcnn folder).
Finally, you will need the data and the trained weights for different parts of the process.
To create the dataset you will need to download the original data and put it in a new /data folder. create_database.py script will patch the images contained in data. On the computer we used (ASUS X756UX) it took approximately 6 hours. Moreover, you should manually move the result of the subdivision in two directories, database/train and database/validate. You should respect the structure created by create_database.py, i.e. patch j from image i for training should be in path database/train/image_i/patch_j. If you don't want to follow the process, we provide a download link to the database.
You will need the database to observe the results of single patch validation, but not to see the final (whole image) results.
For the training, convert the jupyter notebook train.ipynb to a script train.py (jupyter nbconvert --to script train.ipynb) and copy the whole folder to the virtual machine (database included). Clone a new implementation of Mask_rcnn in the virtual machine (the one in the setup guide has been modified) and install it while running the docker image (python setup.py install in Mask_rcnn folder). Run the script on the virtual machine. Weights will be saved in thisfolder/logs/Config1/mask_rcnn_slices.h5 and tensorboard logs will be saved in a new folder in Config1. To visualize the training curves, download from the machine the tensorboard logs and launch tensorboard.
You will need the weights to run validation and whole image results. If you don't want to follow the process, we provide a link to the trained weights. Download and unzip the logs directory.
Download the weights from the machine/from the link into logs/Config1 and run the validation.ipynb notebook. For final whole image results, run the whole_image.ipynb notebook (our "run.ipynb"). Please observed that the image it runs on has not been part of the training dataset of the model.
Original data: https://mega.nz/#!SrIhzICT!sDfA8Oh8wVeRL8R0Tn-TPTDg8BV2PsMauSWf3BIDnKY
Prepared database: https://mega.nz/#!myZDQCba!fH57F_58JYrFqUv_ZqlRBUzarw5FlDEssiUjYqXplrw
Weights and logs: https://mega.nz/#!a7A3EA5T!2X-mWmYderRLs3Bu6beROQlJk7ZYpL_0UbUCnOIlPIo
Final images: https://mega.nz/#!6vJjWCJC!7WYgY5MJyB4p3ARCIS3NUqkNuoM3KqcGOgBgBDYy3p4