Automated PD-L1 Scoring

These files are part of a scientific paper and describe a workflow for automated PD-L1 assessment of Tumor Proportion Score (TPS), Combined Positive Score (CPS), and Immune Cell Score (ICS) with QuPath and MATLAB exemplified with head and neck squamous cell carcinoma (HNSCC).

Puladi, B.; Ooms, M.; Kintsler, S.; Houschyar, K.S.; Steib, F.; Modabber, A.; Hölzle, F.; Knüchel-Clarke, R.; Braunschweig, T. Automated PD-L1 Scoring Using Artificial Intelligence in Head and Neck Squamous Cell Carcinoma. Cancers 2021, 13, 4409. https://doi.org/10.3390/cancers13174409

NOT FOR DIAGNOSTIC USE!

Collaboration

We are open for a further development. Feel free to contact us.

Training data

HNSCC cases from the National Cancer Institute Clinical Proteomic Tumor Analysis Consortium (CPTAC) were used to train the following models.

• Training data can be exported class by class from annotations in QuPath as tile images: exportTiles.groovy
• From HE training data, the hematoxylin channel can be extracted using MATLAB script: deconvolveTiles.m

Pre-trained models / Training new models

Train models for automated PD-L1 scoring

1. First neural network:
   • The first neural network is designed to annotate the tumor. For this purpose shufflenet [1] was used and trained on 8 tissue classes using MATLAB: trainFirstNetwork.m
   • A pre-trained model "shufflenet-HNSCC-NumClass-8.mat" for HNSCC can be found under the folder models.
2. Second neural network:
   • The second neural network is designed to detect the cells.
   • StarDist [2] is used for cell recognition. A pre-trained model can be used for this purpose: he_heavy_augment.zip
3. Third neural network:
   • The third neural network is designed to classify cells into tumor, immune and stromal cells.
   • Training tiles and test tiles must be created for the third neural network. This can be done with the following script in MATLAB: mergeTiles.m
   • Afterwards the merged tile images can be imported into QuPath and the cells can be recognized using the following script: step2_detectCells.groovy
   • By means of the function "Classify → Object classification → Train object classifier" a new classifier can be trained in QuPath.
   • The trained classifier can then be evaluated on another dataset with merged tile images. A helpful Groovy script for this: validateMLP.groovy
   • A pre-trained model "HNSCC_MLP_LYM_TUM_STR.json" for HNSCC can be found under the folder models.

Workflow

1. Annotate tumor:
   • Create labelmap using pre-trained model in MATLAB: step1_1_annotateTumor.m
   • Create project and import WSIs + models (see dependency + models) + import labelmap
   • Prepare WSI with the following script: step_1_2_estimateBackgroundValues.groovy
   • Import labelmap as an annotation into QuPath with the following script: step_1_3_importBinaryLabelmaps.groovy
2. Detect cells:
   • Detect all cells inside annotations with the following script in QuPath: step2_detectCells.groovy
3. Classify cells:
   • Run classifier with the following script in QuPath: step3_classifyCells.groovy
4. Calculate PD-L1 scores:
   • Calculate PD-L1 Scores with the following script in QuPath: step4_calculateScores.groovy

Used software

• QuPath version 0.2.3
• MATLAB 2021a

License

GNU General Public License v2.0

Reference

1. Zhang, X.; Zhou, X.; Lin, M.; Sun, J. ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices, 2017. Available online: https://arxiv.org/pdf/1707.01083.
2. Schmidt, U.; Weigert, M.; Broaddus, C.; Myers, G. Cell Detection with Star-convex Polygons 2018, 11071, 265–273, doi:10.1007/978-3-030-00934-2_30.
3. Bankhead, P.; Loughrey, M.B.; Fernández, J.A.; Dombrowski, Y.; McArt, D.G.; Dunne, P.D.; McQuaid, S.; Gray, R.T.; Murray, L.J.; Coleman, H.G.; et al. QuPath: Open source software for digital pathology image analysis. Sci. Rep. 2017, 7, 16878, doi:10.1038/s41598-017-17204-5.
4. The MathWorks, Inc. Deep Learning Toolbox. Available online: https://www.mathworks.com/help/deeplearning/