Glioma Grading using Machine Learning techniques: Model optimization and web deployment

Detecting and grading glioma at an early stage to determine a tumor's severity is an important step in the treatment of this brain tumor. Although a lot of research on this subject has been based on the use of MRI images in the past few years, molecular markers have grown in their significance in tumor classification. Machine Learning(ML) algorithms have been proven to be very effective in solving problems in the Healthcare sector. This work aims to assess the performance of several ML methods for glioma grading. These ML models' hyperparameters are tuned using the framework Optuna. Using the publicly available TCGA dataset, six ML algorithms including Extreme Gradient Boosting (XGBoost), Categorical Boosting (CatBoost), Ridge Classifier(Ridge), Light Gradient Boosting Machine (LGBM), ExtraTreeClassifier (ExtraT) and Random Forest (RF). We also evaluate the carbon emission produced by each of these models during the training which is one of the important factors to consider when choosing the best model. After comparison of the models using six evaluation metrics such as accuracy, precision, recall, F1-score, AUC, and specificity, XGBoost emerged as the most performant technique with 89.27%, 85.64%, 93.44%, 91.49%, 88.08%, and 91.81% of F1-score, recall, precision, specificity, accuracy, and AUC respectively. We make our model accessible by designing a Web application using the framework Streamlit for real-time predictions and data collection.