Skip to content

Latest commit

 

History

History
58 lines (40 loc) · 2.76 KB

README.md

File metadata and controls

58 lines (40 loc) · 2.76 KB
Raw

oct-irf-train

Retrain our published IRF segmentation algorithm with your own data.

Installation and setup

git clone https://github.com/uw-biomedical-ml/oct-irf-train.git

pip install -r requirements.txt

./run.py <image folder> <csv file>

For example if you added a folder called "data" inside the cloned project folder, and a file called segmentations.csv which should look like:

ptid,original,mask
038883,038883-0.png,038883-0_mask.png
027134,027134-0.png,027134-0_mask.png
029719,029719-0.png,029719-0_mask.png
053971,053971-0.png,053971-0_mask.png
029058,029058-0.png,029058-0_mask.png

Then you would run the command ./run.py data segmentations.csv

This will check your environment and install any packages that are missing. It will also setup the data files into two sets of images: 80% for training and 20% for validation. The files will be packaged and then stored as Numpy arrays inside prepdata/nyparrays. Training will beginning using transfer learning and process for 10 epochs.

The format of the images are expected to be as follows:

  • original OCT B scans = 8-bit grayscale images, ideally PNG.
  • segmentation masks = 8-bit grayscale images with 255 for the feature of interest and 0 for the background.

Output

After training is finished, there will be a folder called run inside the project folder. This folder will also contain the following files:

  • learning.html = This is a HTML file that you should open using a web browser (ideally Chrome) which will show the learning curves.
  • history.txt = This is a text file for the training and validation losses. This is used to generate the learning curves.
  • options.json = The options that were set when this run was initiated.
  • ptgrps.csv = This is how the patients were split into training and validation.
  • weights.hdf5 = This is the weights file for the U-net with the best validation loss during training.

Running your own test set to evaluate the new retrained model

This repositiory also contains an evaluation program which can be run using the following syntax:

./evaluate.py <img folder> <out folder> <mode>
  • img folder = this is where the new images from the test set should be placed. The script will batch process all the images in this folder.
  • out folder = this folder is where to save the resulting images. This folder must be an empty folder and not contain any files inside of it.
  • mode = This should be one of three options:
    • mask = only output the binary mask. This is useful for calculating Dice and IOU metrics.
    • blend = Alpha-blended mask on top of the original OCT B scan image
    • mask_blend = Recommended Using a cutoff of 0.5, a binary mask will be applied and the resulting probabilities will be shown as a color graident overlayed on top of the original B-scan.