mcmasterai / radiology-and-ai Goto Github PK

-Use segementations and images sent by Cameron

-Plan how project will be organized
-Placeholders for where results of future tasks will be

-Wandb and Tensorflow both currently do not allow for this

• Try running wandb logging in online mode

• https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6583803/

• add learning rate parameter to lightning module
• Add wandb logger
• Try to determin how to add hyperparameter search with pytorch lightning
• Save pre-created eval and train dataset
• Update metric logging

-compare results with both Z-score and Nyul normalization
-Use power-law transformation
-Report results

• Seperate augment methods from the collator, 50% (look for what other have done) chance of each augmentation
• Add parameter to lightning module for channels, should set input size accordingly
• Add visualization of segmentation/ make better
• Try to alter lightning module to allow for TPU training should allow for faster training
• Review how output come from medzoo UNET implementation

-Use learning_rate_finding notebook to see which learning rate gives lowest learning rate after 2 epochs
-The notebook is under the training-improvments-47 branch
-The only changes you should have to make are in sections 3 and 4 of the notebook, and in the first cell change
"!git clone https://github.com/McMasterAI/Radiology-and-AI.git" to "!git clone --branch training-improvements-47 https://github.com/McMasterAI/Radiology-and-AI.git"
-Find which accumulate_grad_batches amount finds best learning rate (maybe try 10, 15 and 20)
-Check 50 different learning rates for each
-Each should give different optimal learning rate
-Of these three optimal learning rates, run model training with each and see which gives lowest loss value after 2 epochs of training
-You will need to create a wandb account to track the model training and also make a shortcut to the macai datasets folder I shared with you in your base drive folder
-You don't need to push any changes to the repo just report the best learning rate you found and anything else you think important

• Write up of how augmentations work
• Try to find existing implementation to use instead of the current one
• Cameron will use to speed up the implementations

-Elastic deformations/augmentation

-Use nibabel package
-Try to download Brats 2018 dataset
-Demonstrate on Google Colab

-We want to collator to perform the normalization's we specify

Report the:
-Model architecture used
-Data augmentation method
-The performance metrics of the model
-Data used if available

Focused on people try to complete Brats challenge or something similar

-Ill let everyone know when I figure out when it will be
-We will talk with her about what we have done

-Look at papers that have used Brats dataset for image segmentation in the past
-Write summary of different methods

-use existing training files to see what needs to be contained in the function
-convert command line arguments to method parameters
-place in root of project

• Tumour progression vs pseudo-progression
• Cancer Type (Brain cancer vs Metastases)
• Immunotherapy response
• Any other types

Look into types of architecture, where they got the data, etc

• Refer to the paper Hamza summarized

• Make sure the images normalize such that the backgrounds of the images are zero, and do the intensity normalization
  -Everything below a certain value after z score normalization is the "zero" value
• Run before training and calculate a z-sore he normalizations for each channel
• Pass in as parameter to collator

• Try training while ignoring swelling/edema
• Maybe try telling model to only predict core

-Read in images and annotated data
-Add sequential layers
-Attempt Cat/Dog classification task first https://www.kaggle.com/uysimty/keras-cnn-dog-or-cat-classification

-If successful move onto 3-d image task
-Use methods created by Matt
-https://github.com/NILAB-UvA/AOMIC.github.io
-Binary classification of sex
-Get prediction output
-Show in Google Colab

-Review literature