Skip to content

Latest commit

 

History

History
74 lines (53 loc) · 2.94 KB

README.md

File metadata and controls

74 lines (53 loc) · 2.94 KB
Raw

PACpAInt: a histology-based deep learning model uncovers the extensive intratumor molecular heterogeneity of pancreatic adenocarcinoma

Code for the models described in the paper "PACpAInt: a histology-based deep learning model uncovers the extensive intratumor molecular heterogeneity of pancreatic adenocarcinoma". The study presents an approach to predict molecular subtypes of Pancreatic Ductal adenocarcinoma from H&E histology slides using deep learning.

"PACpAInt"

Four machine learning models are described in the aforementioned paper:

  • PACpAInt-Neo: predict tumor/non-tumor regions (tile-level).
  • PACpAInt-Cell Type: predict tumor cells/stroma regions (tile-level).
  • PACpAInt-BC: predict basal/classic molecular subtypes (slide-level).
  • PACpAInt-Comp: predict basal/classic/stroma active/stroma inactive molecular components (slide-level).

Install

To install openslide, do:

apt-get update -qq && apt-get install openslide-tools libgeos-dev -y 2>&1

Then to install pacpaint and its dependencies:

pip install -e .

Dataset

The general class to load data is PACpAInt dataset present in pacpaint/data/dataset.py. To make it work on your data, some methods should be implemented (see file).

Feature Extraction

ResNet50 pre-trained with supervised learning on ImageNet dataset

In the file pacpaint/engine/feature_extraction/imagenet_extraction.py, we provide an example on how to extract features from each tiles, given their coordinates, using a ResNet50 pre-trained on ImageNet dataset.

Wide ResNet50 x2 pre-training with MoCo-v2 on histology images

The feature extractor used in our study is a Wide ResNet50 x2, that was pre-trained with MoCo v2 on 4 million tiles from TCGA-COAD dataset.

The code to train such model is available here: https://github.com/facebookresearch/moco.

Details about the parameters used for the training are given in our paper.

Predictive models training

Predict tumor/non-tumor (PACpAInt-Neo) and tumor cell/stroma (PACpAInt-Cell Type)

The script to train models to predict tumor and non-tumor regions or tumor cell / stroma regions in a cross-validated fashion is available here: pacpaint/engine/pacpaint_neo_cell_type/train.py.

It can be launched like this:

python pacpaint/engine/pacpaint_neo_cell_type/train.py --model neo

Predict basal/classic molecular subtypes (PACpAInt-BC)

The script to train models to predict basal/classic molecular subtypes in a cross-validated fashion is available here: pacpaint/engine/pacpaint_bc/train.py.

It can be launched like this:

python pacpaint/engine/pacpaint_bc/train.py

Predict basal/classic/stroma activ/stroma inactive molecular components (PACpAInt-BC)

The script to train models to predict the four molecular components in a cross-validated fashion is available here: pacpaint/engine/pacpaint_comp/train.py.

It can be launched like this:

python pacpaint/engine/pacpaint_comp/train.py