ImagePseudo: Image-based inference of tumor cell trajectories enables large-scale cancer progression analysis

Author: Yang Liu (Yang.Liu2@utsouthwestern.edu)

Introduction

An AI-powered approach transforms routine pathology slides into dynamic maps of tumor evolution, enabling large-scale analysis of cancer progression without the need for costly single-cell sequencing.

Citation

Liu Y, Cai L, Rong R, Wang S, Jia L, Quan P, Zhou Q, Xiao G, Xie Y. Image-based inference of tumor cell trajectories enables large-scale cancer progression analysis. Sci Adv. 2025 Jul 18;11(29):eadv9466. doi: 10.1126/sciadv.adv9466. Epub 2025 Jul 18. PMID: 40680117; PMCID: PMC12273761. Pubmed

Overview of ImagePseudo

ImagePseudo Inference

From Input to Output

This repository includes

• Feature extraction from pretained model and tumor grade prediction
• Estimating pseudotime from extracted features
• Inference tumor growth speed, diversity and tumor progression fitness

The support pathology image type

The pathology image types as input.

• svs
• tif
• tiff
• dicom
• ndpi

You can put the pathology image files into one folder. Use the folder path as the input for this project.

Dependencies

• numpy==1.26.4
• pandas==2.2.1
• torch
• torchvision
• transformers==4.38.2
• scanpy==1.9.8
• opencv-python==4.9.0.80
• openslide-python
• openslide-bin
• lifelines==0.28.0
• scikit-learn==1.4.1
• scikit-image==0.22.0
• scipy==1.12.0
• matplotlib==3.8.3
• seaborn==0.13.2
• anndata==0.10.6
• numexpr==2.9.0

You can install them to your conda environment by

pip install requirement.txt

• openslide
• igraph
• leidenalg

Your can install them into your conda environment by

conda install -c conda-forge openslide
conda install -c conda-forge python-igraph
conda install -c conda-forge leidenalg

Build up conda environment

• Install Conda https://docs.conda.io/projects/conda/en/latest/user-guide/install/index.html

You can skip the installation of conda if already installed

• Create conda environment 'imagepseudo' using requirement.txt

conda create --name imagepseudo python=3.9
conda activate imagepseudo
pip install -r requirement.txt
conda install -c conda-forge openslide
conda install -c conda-forge python-igraph
conda install -c conda-forge leidenalg

Download pretrained model and example pathology images in svs format

• Example data and pretrained model are available at Zenode: https://zenodo.org/records/16376974
• [option] Choose Version v2 on right-side link if v1 is chosen
• Download "example.svs" and "model.tar.gz" to this project directory
• Unzip "model.tar.gz" under this project directory, and move to the subfolder "model" (create this folder if not exists)

tar -xvf model.tar.gz
mkdir model
mv model.pt model/

User Guideline

Step 1: Extracting features for tumor grading prediction from WSIs

Input

Input a folder containing all SVS files to proceed.

mkdir input_folder  # create a folder
mv example.svs input_folder/   # move svs file under it. Can be many
python GradePrediction.py --data_path ./input_folder --model model/model.pt --output_path feature/

Input Parameters:

Parameter	Description	Default Value
data_path	single image file path or a folder path containing image files, support svs, tiff, tif, dicom, and ndpi
model	the path of mdoel file 'model.pt'
pixel_step	pixel step for patch extraction	400
output_path	the relative path to output the feature

Output files

• example.svs_286cols_210rows.csv

• example.svs_286cols_210rows.png

Step 2: Exstimate pseudotime from extracted features

python Pseudotime.py --feature_path feature/ --output_path pseudotime/

Input Foler (the folder 'feature' from step 1)

The code will scan the folder 'feature' to deal with all CSV files

• E.g. example.svs_286cols_210rows.csv

Output file

• example.svs_286cols_210rows.csv.h5ad

Step 3: Inference tumor growth speed, diversity and tumor progression fitness

python Fitness.py --input_path pseudotime/ --output_file results.csv

Input Foler (the folder 'pseudotime' from step 2)

The code will scan the folder 'pseudotime' to deal with all h5ad files

• E.g. example.svs_286cols_210rows.csv.h5ad

Output File

• results.csv

TroubleShooting

License

Following UT Southwestern Office for Technology Development, the project is using the license from The University of Texas Southwestern Medical Center.