GeoFed-Cervix: A Differential Geometry–Guided Federated and Explainable AI Framework for Early Cervical Cancer Detection on Consumer Devices

Abstract

Early and accurate detection of cervical cancer (CC) is vital for improving clinical outcomes and guiding timely treatment. We introduce GeoFed-Cervix, a framework with a lightweight Android application (app) designed for multimodal CC diagnostics. The app integrates two modules: GeoFed-CervixYOLO for early CC prediction and GeoFed-CervixLangChain for explainable diagnostics. The framework addresses key challenges in CC screening, including data privacy, model interpretability, and deployment on resource-constrained edge devices. GeoFed-CervixYOLO uses differential geometric priors and personalized federated learning (PFL), ensuring robust feature extraction with privacy-preserving decentralization. GeoFed-CervixLangChain employs LangChain-driven large language models (LLMs) to generate clinician- and patient-specific explanations aligned with explainable AI (XAI) 2.0 principles. Designed for consumer-grade edge devices, the system supports accessibility in low-resource settings. Evaluated on a multimodal dataset of normal and precancerous samples, GeoFed-Cervix achieved 98.27% precision, 98.28% recall, 99.57% specificity, and 98.27% accuracy, while providing interpretable, clinically meaningful insights.

Ähnliche Arbeiten