LUPI

Paper: Learning using Privileged Information for Chemosensitivity Prediction in Ovarian Cancer Patients

Introduction

This project utilizes Whole Slide Images (WSIs) and gene expression data to predict the chemosensitivity in ovarian cancer patients. WSIs can be downloaded from GDC Data Portal. The IDs of each WSI used in this project are available in the data_files/imgs.csv file.

To train and validate the models, one requires tiles extracted from WSIs, gene expression profiles, and their respective labels for each patient.

1. Pre-processing

WSIs are large in size, making it difficult to process the entire image. Therefore, we extracted the top 3 tiles from each slide. Preprocessing code can be found here. This tutorial contains three files: slide.py, filter.py, and tiles.py.

Before executing slide.py, add the path to the WSIs directory and the directory where you want to save the tiles on line 32, 143, and 751. This file will create low-resolution images of each whole slide image. No changes are required in filter.py. In the tiles.py file, change the size of the tile to 1536x2048 and top tiles to 3 on lines 38, 39, and 40, respectively. After these modifications, execute the code to generate the top 3 tiles from each WSI. The sequence is slide.py -> filter.py -> tiles.py.

2. Stain Normalization

Tiles extracted from WSIs need to be stain normalized to remove color variations. To do this, update the local paths in normalization.py and execute it.

After the above two steps, normalized top three tiles for each patient will be obtained. Add these tile paths to the imgs.csv file.

3. Data Files

This folder contains the following three files:

I. imgs.csv

II. genes.csv

III. labels.csv

imgs.csv:

This file contains the path to tiles for each patient used in this project and needs to be updated as per your local path.

genes.csv:

In this file, each column represents a patient’s gene profile, which is already preprocessed as discussed in the paper.

labels.csv:

This file contains labels (-1, 1) for each patient. -1 represents chemo-resistant and +1 represents chemo-sensitive patients.

NOTE: genes.csv and labels.csv files have data in the same sequence as imgs.csv file, which have patient IDs that will be used later for training. Also, change the local path to the tiles directory in imgs.csv file as needed.

Training and Validation

By now, you will have all the data required to train and validate the models. Each model will use the imgs.csv, genes.csv, labels.csv, and loader.py files to load the data, and patch_extractor.py to extract patches from tiles. Make sure you update the local paths in the imgs.csv file.

To train the model, simply execute the model file, e.g., privileged_model.py, which will load the gene profiles of each patient from genes.csv and labels from labels.csv for training and validation of the privileged space model. Also, one can use the testing.py file to utilize the already trained model.