A Hybrid Deep Learning Framework for Early Detection of Ovarian Cancer Using Ultrasound and MRI Images on a Secure Cloud Platform

Ovarian cancer continues to pose a major diagnostic challenge, as early-stage disease often presents with subtle and heterogeneous imaging characteristics that limit the effectiveness of single-modality analysis. In response to this challenge, this study proposes a novel hybrid deep learning framework for the early detection and classification of ovarian cancer using ultrasound and MRI imaging, designed for deployment on a secure cloud-based platform. The proposed framework departs from conventional convolutional and attention-driven models by introducing a capsule-based representation learning strategy that explicitly encodes lesion morphology and spatial hierarchies, enabling robust characterization of small and irregular tumor structures. To capture global contextual relationships across imaging regions without reliance on self-attention mechanisms, the framework integrates an attention-free token-mixing architecture, facilitating efficient long-range interaction while maintaining scalability. In addition, a hypergraph-based relational learning module is employed to model higher-order spatial and radiomic relationships among multiple lesion regions simultaneously, providing lesion-centric reasoning that aligns with clinical diagnostic practices. This combination allows the model to effectively distinguish malignant patterns from benign anatomical variations. Beyond binary cancer detection, the framework supports hierarchical classification, separating benign and malignant cases and further categorizing malignant tumors into clinically meaningful subtypes. To enhance fine-grained discrimination, pathology-guided semantic alignment is incorporated using histopathological knowledge as auxiliary supervision, enabling cross-modal knowledge transfer without the need for paired imaging–pathology data. The framework is evaluated on multiple publicly available datasets covering ultrasound, MRI/CT, and histopathology modalities, demonstrating consistent performance across heterogeneous data sources. To ensure suitability for real-world clinical use, an advanced chaos-based image encryption and secure transmission module is integrated to protect sensitive medical data during cloud-based processing. Experimental results indicate that the proposed framework achieves superior detection and classification performance compared to existing approaches, particularly in early-stage ovarian cancer cases, underscoring its potential as an accurate, interpretable, and clinically deployable decision-support system.