Brain Cancer Segmentation and Diagnostic Platform Using Deep Learning and Federated Learning

Abstract

This thesis addresses the critical challenge of accurate brain cancer diagnosis while ensuring patient data privacy through the integration of deep learning (DL) and federated learning (FL). A novel 3D U- Net architecture with adaptive pooling techniques is developed, achieving superior tumour segmentation performance (Dice similarity coefficient >0.89) on BraTS datasets by effectively capturing heterogeneous tumour features. The model incorporates a Choquet integral-based fusion mechanism to enhance diagnostic precision by modeling complex feature interactions. To overcome data privacy barriers, a differentially private FL framework is implemented, enabling secure multi- institutional collaboration without raw data sharing. This approach maintains segmentation accuracy (DSC: 0.86–0.91) while complying with GDPR/HIPAA regulations through secure aggregation and differential privacy. The research culminates in a diagnostic platform prototype that integrates segmentation, visualization, and treatment planning tools, validated for clinical utility. Key contributions include: (1) State-of-the-art segmentation performance via optimized 3D pooling and Choquet fusion; (2) Privacy-preserving FL that matches centralized model accuracy; (3) An end- to-end platform supporting real-time clinician collaboration. Future work will extend this framework to broader oncological pathologies and enhance model interpretability for clinical adoption.