ExpertDrivenNuclei

Implementation of "An Expert-driven Data Generation Pipeline for Histological Images".

This repo uses AdaIN and HoVerNet in order to complete the pipeline and evaluate the generated datasets.

Replicating the generation pipeline

Preparing the datasets

To prepare the public data, download the datasets described in public_datasets_selection.ipynb, unzip them and run the notebook changing the base_dir path.

Once a unified representation is obtained, run the masks_to_HoVerNet.py script to convert all annotations to the format required for training HoVerNet. split_dataset.py can be then used to split the dataset in the given subsets.

Training AdaIN

To train the style transfer, select a subset of images associated to their flattened ground truths:

Follow the instructions available here, selecting as content_dir the one containing the flattened masks and as style_dir the one with real images.

Our pipeline uses the last decoder checkpoint.

Generating synthetic datasets

A new dataset can be generated by running generate_dataset.py. The script requires as parameters:

• real_path: the path to the folder containing real images in HoVerNet format (i.e., the folder should contain an Images and a Labels folders)
• adain_path: the path to the folder containing the AdaIN repo (for calling the inference script)
• adain_decoder_path: the path to the trained decoder (e.g. decoder_iter_160000.pth.tar)
• styles_path: folder containing target styles for AdaIN

For a complete list of available parameters, launch the script with the -h option.

HoVerNet training and evaluation

New datasets are already generated in a format compatible with HoVerNet. To train the segmentation network, follow HoVerNet's instructions. Set also in config.py:

• nr_type = None
• self.type_classification = False
• self.dataset_name = "cpm17" to use the correct annotation format and perform only segmentation.

HoVerNet evaluation was performed with the following metrics (already available in the framework):

• DICE
• DICE2
• AJI
• AJI+