Tumor Recurrence Prediction After Stereotactic Radiosurgery: Multimodal Machine Learning Approaches in Imbalanced Data Contexts

Gamma Knife is an advanced form of radiation therapy allowing non-invasive tumor treatment by delivering high radiations doses to localized brain regions. Early detection of tumor recurrence after Gamma Knife radiation therapy can significantly influence treatment decisions and improve outcomes in brain cancer patients. In this work we propose several machine learning solutions to predict local recurrence of metastatic brain tumors after Gamma Knife radiation therapy exploiting MRI scans, radiation dose maps and clinical patient information. The training set consists of 140 stable and 13 recurrent lesions, while the test set contains 81 stable and 10 recurrent lesions. Oversampling, focal loss, data augmentation and thresholding are adopted to face data imbalance along with cross validation to improve model generalization. We explored different multimodal fusion techniques: input fusion via Discrete Wavelet Transform, concatenation-based information fusion, LSTM-based information fusion and attention-based fusion. Despite the challenges posed by data scarcity and imbalance, our proposed solutions delivered promising results proving the potential of multimodal machine learning models leveraging MRI, Dose and clinical information, laying the foundations for future studies aimed at timely, accurate tumor recurrence predictions to improve life quality and expectancy of patients affected by metastatic brain cancer.