A privacy-preserving cloud-edge collaborative learning model for uterine ultrasound image analysis

Abstract

This invention discloses a privacy-preserving cloud-edge collaborative learning model for uterine ultrasound image analysis, belonging to the field of medical image processing technology. It constructs a dataset and divides it into cloud-based public data and edge-based private data. Employing a collaborative learning strategy of "cloud pre-training + edge fine-tuning," a semantically enhanced visual-language model is built in the cloud, using a clinical knowledge graph to construct a semantic similarity alignment mechanism for pre-training. At the edge, each client performs federated fine-tuning based on private data, using semantic matching loss to constrain local updates and resist model drift caused by non-independent and identically distributed data. Simultaneously, a differential privacy mechanism is introduced in the local model update, using gradient pruning and noise injection to prevent privacy leakage. In the context of strict data privacy requirements for consumer electronic devices, this invention effectively solves the problems of data silos and limited resources, achieving high-precision image processing comparable to centralized training while ensuring data privacy.

Images