- WSI pre-processing tools are available at https://github.com/IARCbioinfo/WSIPreprocessing
- This repository contains:
- Scripts for dividing WSI into patches, called tiles.
- Scripts to normalize the HE/HES coloring of WSIs, used to artificially remove saffron coloring from WSIs produced in the French center.
Tumor segmentation with CFlow AD:
- Tumour areas were segmented unsupervised using the CFLow anomaly detection model. An adaptation of this model for this task is available at https://github.com/IARCbioinfo/TumorSegmentationCFlowAD
- This repository contains: + Scripts to train and evaluate the model + Script to create the segmentation maps
- Training sets for Ki-67 and HE/HES WSI are available on request from mathiane[at]iarc[dot]who[int], as is the pre-trained model (will be available on a server soon).
Automatic assessment of lung neuroendocrine neoplasms (LNEN) proliferative activity using Pathonet
- The Ki-67 and PHH3 indices were quantified automatically using the Pathonet supervised deep learning model. An adaptation of this model is available at the following address https://github.com/IARCbioinfo/PathonetLNEN
- This repository also contains scripts for calculating spatial metrics using graph theory as proposed by Bullloni and colleagues See : Automated analysis of proliferating cells spatial organization predicts prognosis in lung neuroendocrine neoplasms, Cancers 2021
- Annotated LNEN tiles and network weights are available on request from mathiane@iarc.who.int.
WSI features extraction using Barlow Twins
- The unsupervised deep learning model called Barlow Twins, proposed by J. Zbontar and colleagues, was used to extract the features of the tiles composing the HE-stained WSIs of LNEN patients. The adaptation of the method to the pathology that we have developed in PyTorch is available at https://github.com/IARCbioinfo/LNENBarlowTwins.
- LNEN pre-processed tiles and network weights are available on request from mathiane@iarc.who.int.
- The projections obtained from the Barlow twins are independent of the location of the tiles on the WSI. To model the fact that tiles spatially close to each other on the WSI are more likely to share common morphological features, we adapted the spatial principal component analysis (PCA) originally proposed for WSIs by L. Shang and X. Zhou, NAT COM 2022.
- The R scripts are available at https://github.com/IARCbioinfo/SpatialPCAForWSIs.
- This repository also contains the method for calculating the Leiden community, which allows tiles that are more likely to have similar morphological features to be grouped together in the same cluster.
- 🚧 Some stats script