RMS_AI

Github page for Predicting survival of rhabdomyosarcoma patients based on deep-learning of hematoxylin and eosin images manuscript

Set Up Environment

Two different conda environments are required

• rms1: 'subtype', 'myod1' and 'tp53' tasks
• rms2: 'segment' and 'survival' tasks

conda env create -f environment1.yml
conda activate rms1
conda install pytorch=1.7.1 torchvision=0.8.2 torchaudio cudatoolkit=10.2 -c pytorch

conda env create -f environment2.yml
conda activate rms2
conda install pytorch=1.7.1 torchvision=0.8.2 torchaudio cudatoolkit=10.2 -c pytorch

Repository Structure

Below are the src subdirectories in the repository:

• src/results/: output files are stored at different subdirectories
• src/images/: each subdirectory contains image files (svs, tif or png) to be inferenced
• src/images/cancer_maps/: contains greyscale images where cancerous regions are defined
• src/model_weights/: trained model weights are located at different subdirectories

Below are the main python files in the repository:

• src/subtype_inference.py: WSI subtype classification
• src/segmentation_inference.py: Pixel-level segmentation of different features in WSIs
• src/tp53_inference.py: TP53 mutation detection
• src/myod1_inference.py: MYOD1 mutation detection
• src/survival_inference1.py and src/survival_inference2.py: Risk prediction
• src/ras_inference.py: RAS mutation detection

Model weights

Trained model weights can be found using the link below

• https://figshare.com/articles/software/segment_fold01/19182074

Please unzip the downloaded file and place the folder under the 'src' directory