Predicting the Local Response of Metastatic Brain Tumor to Gamma Knife Radiosurgery by Radiomics With a Machine
Daisuke Kawahara1, Xueyan Tang2, Chung K Lee2
1Department of Radiation Oncology, Institute of Biomedical & Health Sciences, Hiroshima University, Hiroshima, Japan.
Frontiers in Oncology
|January 28, 2021
Summary
This study developed a machine learning model using radiomics to predict brain metastases response to Gamma Knife Radiosurgery (GKRS). The model significantly improves prediction accuracy compared to visual evaluation, aiding clinical decisions.
Area of Science:
- Radiology
- Oncology
- Medical Imaging
- Machine Learning
Background:
- Brain metastases (BMs) pose a significant challenge in cancer treatment.
- Predicting treatment response is crucial for optimizing patient management.
- Gamma Knife Radiosurgery (GKRS) is a common treatment modality for BMs.
Purpose of the Study:
- To develop and validate a machine learning (ML) model for predicting the local response (LR) of BMs treated with GKRS.
- To utilize radiomics features extracted from MRI to enhance prediction accuracy.
- To provide a decision support tool for clinicians to optimize treatment outcomes.
Main Methods:
- A cohort of 157 BMs treated with GKRS was analyzed using contrast-enhanced T1-weighted MRI.
- Radiomics analysis extracted over 700 features, with feature selection performed using LASSO regression.
- A neural network (NN) classifier with 10 hidden layers was trained and validated using five-fold cross-validation.
Main Results:
- Seven radiomics features were identified as significant predictors of BM response.
- The developed NN model achieved an accuracy of 78% and sensitivity of 87%, with an AUC of 0.87.
- This performance significantly outperformed the traditional visual evaluation method (accuracy 44%, sensitivity 54%).
Conclusions:
- The proposed NN model, leveraging radiomics features, offers a more accurate and reliable method for predicting BM response to GKRS.
- This ML-driven approach can enhance clinical decision-making and improve patient management strategies.
- Radiomics-based prediction models hold significant promise for personalized cancer treatment planning.
