Secure Cloud-Based Hybrid Deep Learning for Early Diagnosis of Ovarian Cancer Using Federated Learning

Abstract

Ovarian cancer is often diagnosed at later stages due to subtle symptoms and complex imaging, resulting in lower survival rates. This study proposes a novel cloud-based deep learning framework for early ovarian cancer detection using the using the UBC-OCEAN histopathology dataset, which consists of Hematoxylin and Eosin (H&E)–stained ovarian cancer tissue images acquired as Whole Slide Images (WSIs) and Tissue Microarrays (TMAs). Our hybrid model combines Convolutional Neural Networks (CNNs), Transformers, and Graph Neural Networks (GNNs) to capture detailed local features, understand long-range relationships, and identify spatial tumor patterns. The CNN focuses on feature extraction, the Transformer captures contextual relationships, and the GNN highlights spatial connections between lesions. This multi-stage approach significantly improves the detection of small and difficult-to-identify malignant patterns in histopathological images. The system operates on a secure cloud platform, ensuring real-time processing of high-resolution medical images. Patient data is protected using AES-256 encryption, TLS for secure data transfer, and two-factor authentication (2FA) for access control. Federated Learning enables collaborative model training across institutions while maintaining data privacy and compliance with HIPAA and GDPR. Our model achieves exceptional performance with a sensitivity of 98.43% and specificity of 98.71%, outperforming existing models like ResNet-50, VGG16, and InceptionV3. Additionally, the security framework blocks 99.49% of unauthorized access, ensuring robust data protection. This research demonstrates the potential of integrating hybrid deep learning methods with secure cloud systems for the early detection of ovarian cancer in clinical settings.