brca_unet_finegrain_segmentation

This repo is for training and running prediction of brca finegrain segmentation using Unet model.

Dependencies

• Pytorch 0.4.0
• Torchvision 0.2.0
• cv2 (3.4.1)
• Openslide 1.1.1
• sklearn
• PIL

Setup data folder

• The training/validation data are saved in the .data folder
• Training patches and corresponding masks are stored in separate folders. Here is an example:
• Training data:
  • TCGA_BRCA_finegrain_patches_10X
  • TCGA_BRCA_finegrain_patches_10X_mask
• Validation data:
  • TCGA_BRCA_finegrain_patches_10X_val
  • TCGA_BRCA_finegrain_patches_10X_mask_val
• Number 10X indicates the resolution in which the patches were extracted. There is an argument in train.py, 'resolution', to select the resolution if you have different resolution settings.
• The training patch and the corresponding mask have the same size, ideally square patch, e.g 1000x1000 pixels.
• The mask contain values 0, 1, 2,... to indicate the class ID. For binary classification, mask's values are 0/1
• Change the data folder paths accordingly in util_codes/utils.py to load training/validation data

Training

• Run python train.py to train the model.
• Arguments can be passed to select input size of the model, batch size, learning rate, etc.

Predict

• Run python predict.py to run prediction for validation patches.
• Predicted patches will be stored in data folder, under the name "predicted_imgs_10X", if the patches are at 10X
• The predicted patches are original patches overlayed with predicted masks.