PrognosAIs_glioma

This repository contains the code and trained models required for the running of the PrognosAIs model trained for the prediction of the IDH mutation status, 1p/19q co-deletion status, and grade of glioma based on pre-operative MRI scans. The model also automatically segments the tumor.

The model and work that went into the development of the model are described in the paper:

Sebastian R van der Voort, Fatih Incekara, Maarten M J Wijnenga, Georgios Kapsas, Renske Gahrmann, Joost W Schouten, Rishi Nandoe Tewarie, Geert J Lycklama, Philip C De Witt Hamer, Roelant S Eijgelaar, Pim J French, Hendrikus J Dubbink, Arnaud J P E Vincent, Wiro J Niessen, Martin J van den Bent, Marion Smits, Stefan Klein, Combined molecular subtyping, grading, and segmentation of glioma using multi-task deep learning, Neuro-Oncology, 2022;, noac166, https://doi.org/10.1093/neuonc/noac166.

Please cite this paper when you use this code.

Running the model

Docker

The model can be run using the pre-built Docker which contains all code and data needed to apply the model and will carry out all the required pre-processing. The docker can be run as follows:

docker run -u $UID:$GROUPS -v "<local_input_folder>:/input/" -v "<local_output_folder>:/output/" svdvoort/prognosais_glioma:1.0.2

Here <local_input_folder> needs to be replaced by the path to the folder on the host machine that contains the scans for the different subjects. <local_output_folder> needs to be replaced by the folder on the host machine in which the results should be stored.

<local_input_folder> should contain one folder per subject, with for each subject the NIFTI files for the pre-constrast T1-weighted scan, post-contrast T1-weighted scan, T2-weighted scan, and T2-weighted FLAIR scan named T1.nii.gz, T1GD.nii.gz, T2.nii.gz and FLAIR.nii.gz respectively.

Thus an example of a structure for two subjects would be:

<local_input_folder>
|
|  Subject-001
|  |  FLAIR.nii.gz
|  |  T1.nii.gz
|  |  T1GD.nii.gz
|  |  T2.nii.gz
|
|  Subject-002
|  |  FLAIR.nii.gz
|  |  T1.nii.gz
|  |  T1GD.nii.gz
|  |  T2.nii.gz

The outputs of the model are then saved in <local_output_folder>/Results. A mask is stored for each patient as <subjec_id>_mask.nii.gz and the results of the genetic and histological feature predictions are all stored in genetic_histological_predictions.csv.

Locally

You can also evaluate the model by locally installing it. This requires you to do your own pre-processing according to the article: https://doi.org/10.1016/j.dib.2021.107191. The easiest way to set-up the pipeline locally is to follow the same steps as provided in the docker file. You can then run the pipeline script to evaluate the model.

Model

If you are just interested in the model it is available in the Trained_models folder.

The model is compressed unto a tar archive. To restore the model:

cd Trained_models
tar -xzvf prognosais_model.tar.gz

The model is now stored,in TensorFlow SavedModel format, in the prognosais_model folder.

FAQ

I get en error like: /run_pipeline.sh: line 51: 276 Illegal instruction when running the docker

If you are trying to run the docker on a newer Mac with an M1/M2/M3 etc. chip this might be the cause. See this issue. The best approach is to run the model locally instead of using Docker in this case. If you run into this error but are not running on a Mac, feel free to open a new issue.

How should I interpret the file with the genetic predictions?

• Prediction_IDH_class_0 is the prediction score for IDH wildtype
• Prediction_IDH_class_1 is the prediction score for IDH mutated
• Prediction_1p19q_class_0 is the prediction score for 1p19q not-co-deleted
• Prediction_1p19q_class_1 is the prediction score for 1p19q co-deleted
• Prediction_grade_class_0 is the prediction score for grade II
• Prediction_grade_class_1 is the prediction score for grade III
• Prediction_grade_class_3 is the prediction score for grade IV