This research is carried out as part of the project AiCOLO and has been been published in the conference IEEE 36th International Symposium on Computer Based Medical Systems (CBMS) 2023. The paper is available here. This work is a deep learning project applied to medical images. This project introduces a methodology to train deep neural networks to classify genetic mutations directly from histopathological images.
- Before executing the scripts, make sure you have correctly edited the configuration file: config.cfg
- Medical images with annotations of tumor areas and genetic mutations (The medical images used in this project come from the AiCOLO private dataset).
- Tensorflow-Keras / Python
- The dataset should be organized as follows:
- If you use QuPath to annotate tumor areas in your medical images, you can use these files to cut patches from the annotated areas:
- python 1_move_patches_from_QuPath.py
- python 2_create_patches_from_QuPath_patches.py
- Apply the staining normalization:
- python 3_normalize_patches.py
- Organize patches into 5 folds for cross-validation:
- python 4_organize_split.py
- Train one network per fold:
- python 5_experiment_train.py
- Create and save network training and evaluation graphs:
- python 6_create_graphs.py
- Draw the predictions for one WSI:
- python 7_draw_preds_on_WSI.py
- Run CAM or LIME to understand network predictions
- python 8_run_explainability.py
Robin Heckenauer - robin.heckenauer@gmail.com
This work was supported by the AiCOLO project funded by INSERM/Plan Cancer. The computational resources were provided by the Mesocentre of the University of Strasbourg