MLP Group 5: Metastatic Cancer Detection

Build a breast cancer metastases detection and classification model using deep learning.

How it works

1. Divide the WSI into fixed-size tiles and label each tile as positive or negative
2. Train a Convolutional Neural Network (CNN) to classify the WSI as tumorous or normal
3. Predict a score (between 0 and 1) for each tile in a WSI randomly picked from the test set
4. Combine prediction results on tiles to create a heatmap which shows the metastases in the WSI

Data

The datasets used in this project: CAMELYON16 and CAMELYON17 are available in the CAMELYON challenge website

The datasets consist of Whole-Slide Images (WSI) of lymph nodes sections.

CAMELYON16

• Contains a total of 400 whole-slide images.
• Training set contains 270 WSI. The ground truth data for the slides containing metastases is provided in two formats: .xml files containing metastasis annotations and WSI binary Masks.
• Test set contains 130 WSI.

CAMELYON17

• Contains a total of 1000 whole-slide images.
• Training set and test set contain each 100 patients, and each patient consists of five WSI.
• Training set also contains 50 slides with detailed annotations of metastases.

Prerequisites

The modules needed to run this project can be installed using pip:

• scipy==1.3.1
• progress==1.5
• graphviz==0.13.2
• openslide==3.4.1
• pandas==0.25.3
• scikit-learn==0.21.3
• scikit-image==0.16.2

Execution Instruction

1. Add the custom_modules directory to PYTHONPATH environment variable

This directory contains the classes and functions neccessary for handling WSI and preprocessing the dataset.

2. Run generate_tiles.py located in preprocess

Divide WSI into fixed-size 256 x 256 tiles, and separate them into positive tiles (using the annotations provided) and negative tiles (using Otsu thresholding).

3. Run generate_hdf5.py located in preprocess

Create a single HDF5 file containing all tiles obtained with generate_tiles.py (this is useful in reducing reading time).

4. Run model.py located in cnn_model

Train a CNN to predict if a tile is tumorous or normal

5. Run heatmap.py located in heatmap_creation

Use the score predicted by the model on individual tiles to construct a heatmap of the metastases in the WSI.

Trained model

The trained model is available here