3D Multi-Input Deep Learning for Brain Lesion Classification: Attention-Based Analysis of Stable vs. Recurrent Lesions

Brain metastases are a common and serious complication among cancer patients and are often treated by stereotactic radiosurgery, such as Gamma Knife therapy. While this treatment does an excellent job of controlling the local tumor, it is necessary to discriminate between stable and recurrent disease. Existing evaluation strategies are largely based on expert visual interpretation of serial MRI and Radiotherapy-Planning Images; however, this approach requires intensive manual handling of Radiotherapy Dose data and, as such, introduces variability across operators and observers. This thesis aims to evaluate the current literature on the topic and introduces an effective deep-learning architecture designed to classify brain lesions as stable or recurrent after Gamma Knife radiosurgery. The resulting model integrates multimodal information such as Magnetic Resonance Imaging (MRI), Radiotherapy Dose distributions (RTDose), and highly structured clinical parameters into a single multi-input neural network architecture. Particular consideration is given to handling the large class imbalance in recurrence prediction through the application of selective augmentation and balanced sampling schemes to enhance learning efficiency. Extensive experimentation and validation demonstrate meaningful improvements over existing baselines, with greater robustness across patients and different data. Compared to previously established metrics (10% recall, 18.2% F1 score), the model achieved a significant improvement with a recall of 50% and an F1 score of 28.6%. This sensitivity improvement is clinically significant and may help avoid critical interventions for recurrent cases from being delayed. The proposed framework contributes meaningfully to the evolving field of automated neuro-oncology, laying the foundation for consistent, data-driven monitoring of patients undergoing radiosurgical treatment for brain metastases. Although the results and findings are encouraging, demonstrating the potential of combining deep learning techniques, multimodal imaging data, and structured clinical information, they also indicate that this fundamental and important topic requires further and greater focus.