Explainable AI and Advanced Mathematical Modeling for Radiotherapy Outcome Assessment

Abstract

Brain metastases (BMs) are the most common intracranial tumors in adults and remain a major cause of morbidity and mortality in patients with systemic cancer. Stereotactic radiosurgery (SRS) has become a standard treatment modality for managing BMs due to its ability to deliver highly conformal, high-dose radiation while sparing surrounding normal brain tissue. Despite its widespread clinical adoption and favorable outcomes, several challenges persist in the management of SRS-treated BMs. These include limited ability to predict local control (LC) outcomes at the time of treatment planning, difficulty differentiating tumor recurrence (TR) from radionecrosis (RN) on post-treatment imaging, and the need for accurate, robust, and generalizable tumor segmentation to support treatment planning and longitudinal monitoring. Addressing these challenges requires computational frameworks that achieve high predictive performance while remaining interpretable, robust to data heterogeneity, and suitable for real-world clinical deployment.This dissertation develops and evaluates a series of artificial intelligence (AI) frameworks that integrate deep learning, advanced mathematical modeling, and explainability-driven design to improve radiotherapy outcome assessment for SRS-treated brain metastases. The work is organized around three interconnected objectives spanning outcome prediction, diagnostic classification, and automated segmentation. The first aim of this dissertation focuses on lesion-level prediction of local control following SRS. A novel deep ensemble learning framework was developed to integrate pre-treatment contrast-enhanced T1-weighted magnetic resonance imaging (T1-CE MRI), patient-specific spatial dose distributions (Dmap), target volume information, and clinical variables. A spherical image projection strategy was introduced to transform planar image content into multiple locoregional representations, enhancing spatial sensitivity to dose heterogeneity and tumor-tissue interactions. Four VGG-19-based encoders with distinct spherical projection configurations were incorporated into an ensemble design, with clinical features fused via positional encoding. The ensemble outputs were combined using logistic regression to produce final predictions. Evaluated on a cohort of 114 BMs from 82 patients, the proposed framework achieved high predictive performance with strong sensitivity, specificity, and area under the receiver operating characteristic curve, outperforming single-network and non-clinical-feature baselines. These results demonstrate that integrating spatial dose information and ensemble learning enables early identification of high-risk lesions at the time of treatment planning. The second aim addresses the post-treatment challenge of differentiating radionecrosis from tumor recurrence, a task that remains clinically difficult and often requires invasive biopsy. To improve diagnostic performance while enhancing model transparency, a radiogenomic framework based on neural ordinary differential equations (NODEs) was developed. By modeling deep feature extraction as a continuous dynamical system governed by a second-order heavy-ball NODE formulation, the evolution of image, genomic, and clinical features was explicitly tracked across network depth. This approach enabled reconstruction of a decision field in latent space, visualization of feature trajectories, and quantitative assessment of feature contributions over virtual time. The model identified key intermediate states that were aggregated for final classification. Evaluated on a cohort of non-small cell lung cancer patients with biopsy-confirmed outcomes, the NODE-based framework achieved strong diagnostic performance while providing dynamic, interpretable insights into the relative roles of clinical, imaging, and genomic features. This work demonstrates the feasibility of embedding explainability directly into deep learning architectures for radiotherapy response assessment. The third aim focuses on automated BM segmentation in multi-institutional settings, where data heterogeneity and privacy constraints limit centralized model training. Two federated learning (FL) frameworks were developed to improve robustness and generalizability without direct data sharing. The first framework combined federated learning with a fisheye-inspired spherical image transformation and ensemble inference strategy. Multiple locoregional views were generated for each three-dimensional MRI volume using inhomogeneous spherical scaling, and predictions were back-projected into a common anatomical space for ensemble fusion. This design improved segmentation performance at data-limited institutions while preserving accuracy at data-rich sites. The second framework extended federated learning by incorporating uncertainty estimation via test-time augmentation directly into the optimization objective. A voxel-wise uncertainty penalty was introduced to suppress unstable predictions within tumor regions, reducing false-positive detections and improving segmentation reliability. Across both frameworks, federated models demonstrated improved performance and robustness compared to locally trained models, while uncertainty scores indicated enhanced calibration and confidence across institutions. Collectively, this dissertation demonstrates that integrating explainable AI principles with advanced mathematical modeling and federated optimization can substantially improve radiotherapy outcome assessment for brain metastases. The proposed methods address key clinical challenges across the full treatment trajectory, from pre-treatment risk stratification to post-treatment diagnosis and longitudinal monitoring. By emphasizing explainability, robustness, and generalizability alongside predictive accuracy, this work establishes a unified framework for clinically meaningful AI deployment in neuro-oncology and provides a foundation for extending these approaches to other radiotherapy applications.