مقالات پذیرفته شده در نهمین کنگره بین المللی زیست پزشکی
Predicting Treatment Response in Vestibular Schwannoma Patients Undergoing Gamma Knife Radiosurgery Using individual and Combined Radiomics and Dosiomics with Machine Learning
Predicting Treatment Response in Vestibular Schwannoma Patients Undergoing Gamma Knife Radiosurgery Using individual and Combined Radiomics and Dosiomics with Machine Learning
Mohadeseh Gholi,1,*Bijan Hashemi,2Hassanali Nedaei,3
1. Department of Medical Physics, Faculty of Medical Sciences, Tarbiat Modares University, Tehran, Iran 2. Department of Medical Physics, Faculty of Medical Sciences, Tarbiat Modares University, Tehran, Iran. 3. Department of Radio oncology, Radiation oncology research center, Tehran University of Medical Sciences, Tehran, Iran
Introduction: Vestibular schwannoma (VS) is the third most common benign brain tumor. By affecting auditory regions, this condition leads to various complications, including tinnitus, hearing loss, vertigo, and headaches, as well as severe cases of hydrocephalus and brain enlargement. Despite being benign, prompt and effective treatment is crucial because this type of brain mass can impair brain function and cause irreversible damage to brain tissue by involving sensitive regions. Predicting treatment response in patients before initiating therapy is highly important to avoid ineffective treatments, reduce side effects, and prescribe personalized treatment protocols based on individual characteristics. Accordingly, this study evaluates the performance of various machine learning models using radiomics and dosomics data to predict treatment response in vestibular schwannoma patients undergoing Gamma Knife radiosurgery.
Methods: This study utilized data from 65 Gamma Knife-treated patients at Yas Hospital, including MRI images (T1 FSPGRD with and without contrast), dose distribution data, patient age and sex, and six-month follow-up outcomes classified into response and progression groups. These patients underwent 3 sessions of treatment with a total dose of 21 Gy. After preprocessing the MRI images, radiomic features were extracted using the Python PyRadiomics library. Four feature selection methods—MRMR, LASSO, RFE, and ANOVA—were applied alongside four machine learning models (Support Vector Machine (SVM), Decision Tree, Random Forest, and XGBoost) using cross-validation. The models were evaluated across four feature categories: non-contrast radiomics, contrast-enhanced radiomics, dosomics, and combined radiomics-dosomics (with patient age and sex included in all categories). Performance metrics including AUC, accuracy, sensitivity
and specificity were calculated to assess model efficacy.
Results: The best-performing models were selected based on the AUC metric. Across all feature categories, the XGBoost (XGB) machine learning model demonstrated superior performance. For both contrast-enhanced and non-contrast radiomics classifications, the XGB model with LASSO feature selection achieved the highest AUC (92%). In contrast-enhanced radiomics, sensitivity (SEN), specificity (SPE), and accuracy (ACC) were 87%, 88%, and 96%, respectively, while non-contrast radiomics yielded 81%, 81%, and 90%. In the dosomics category, XGB with ANOVA feature selection performed best, achieving an AUC of 93% with SEN, SPE, and ACC values of 86%, 81%, and 96%, respectively. The combined radiomics-dosomics classification using XGB with LASSO and ANOVA feature selection achieved the highest overall performance (AUC: 96%; SEN: 97%; SPE: 95%; ACC: 98%), outperforming all other models and categories.
Conclusion: This study demonstrated that machine learning models based on radiomics and dosomics data can effectively predict treatment response in vestibular schwannoma patients. The XGBoost model consistently outperformed others across all categories, with the combined radiomics-dosomics approach (using LASSO feature selection) yielding the best results. Dosomics-based models surpassed radiomics-only models, while non-contrast radiomics exhibited the weakest performance.