This repository offers an enhancement to the article:

Smart Learning of Click and Refine for Nuclei Segmentation on Histology Images

1. Download PanNuke Dataset:

You can Download the dataset here:PanNuke.

The dataset contains 3 folders: folder 1,2,3.

In the config file config, precise the folder location:

path_panuke = FOLDER_LOCATION

2. Process data:

run

python np_to_images_folder.py

this code:

• 1. $~~~~~$ Opens the images.npy and mask.npy for each folder
• 2. $~~~~~$ Corrects wrong annotations
• 3. $~~~~~$ Saves each patch and each ground truth annotations as .tif files

3. Run train/test/val split:

run

python train_test_val.py --ptrain 0.75  --pval 0.125

This code generates 3 random dataframes. Each dataframe contains filenames of the images belonging to the correponding split. You can choose p_train, p_val for the proportion of each split.(p_test = 1 - p_train-p_val).

4. Train Autoencoder:

run

   cd Autoencoder
   python train.py

5. Generate Contour/Inside from baseline

run

python generate_contour_inside.py --path_annotations

path_annotations is the path of the annotations. It can be the ground truth annotations or the predictions from a baseline nuclei segmentation on your images.

If you have a baseline nuclei segmentation of your images, store it in path_baseline\baseline.

To convert the ground truth annotations to contours and masks choose path_annotations=path_gt To convert the baseline predictions to contours and mask choose path_annotations=path_baseline

6. Modify the baseline segmentation by adding merged and splited nucleis.

run

python fast_augment_save.py

If the baseline nuclei segmentation doesn't have much splits/merges.

This code create an augmented segmentation by adding split and merges errors on random nucleis in each images and save the new annotations, contours and insides in a new folder.

It also creates a clickmap for each image by comparing the ground truth to the baseline segmentation.

The click map has 4 channels:

• 1^st channel is for False Positive nuclei (FP)
• 2^nd channel is for merged nucleis
• 3^rd channel is for splitted nucleis
• 4^th channel is for FN nuclei (FN)

7. Train Click_ref

cd Click_ref
python train.py

This code uses the baseline segmentation and the click maps generated to train the click_ref module to reconstruct the ground truth.

8.