Game theoretic interpretability for learning based preoperative gliomas grading

Abstract Gliomas are the most common primary tumors occurring in the central nervous system. Accurate gliomas grading is crucial for prognosis assessment and optimal treatment on the part of the patient. This study aims to develop and validate a pretreatment MRI-based noninvasive machine learning radiomics model for preoperatively grading glioma, and to simultaneously interpret the radiomics features used in the model. Firstly, wavelet transform and Laplacian of Gaussian (LoG) filtering are used during image preprocessing and a total of 1024 quantitative features are extracted from the region of interest (ROI) of the tumor, which is manually delineated on the largest slice of MRI images. Then, feature selection is performed by Pearson correlation coefficient and the least absolute shrinkage and selection operator (LASSO). Finally, extreme gradient boosting (XGBoost) is built to carry out the glioma grading, and Shapley value is used to quantitatively interpret and reveal the important features contributing to grading. Experimental results on a benchmark dataset demonstrate that XGBoost is effective and efficient for the diagnosis of glioma. Accuracy, sensitivity, specificity, and AUC are 0.83, 0.86, 0.81, 0.86, respectively. These results add evidence of the important role of radiomics model based on only one representative MRI image in preoperatively grading glioma. The quantitative analysis and interpretation may assist clinicians to better understand the disease and select appropriate treatment for improving clinical outcomes.