With the rapid growth of the Internet of Things (IoT) and 6G, an exponential growth of the data will be generated with the gateway. Considering the bandwidth and computation resource requirement of the IOT, the Edge-Cloud collaborative framework is developed to shrink the latency and overloading. With respect to Edge-Cloud collaborative scene, privacy protection, and personalized edge training is necessary. For example, in the Internet of Medical Things area, pathological data of patients cannot be shared from one hospital to another hospital without permission. And respect to the unknown disease, due to little knowledge about the pathogenesis, plenty of pathological data cannot be annotated. In this paper, a Semi-supervised Federated learning-based Edge-Cloud Collaborative Framework, namely Fedsemgan, which leverages Generative Adversarial Networks (GANs) as a semi-supervised Federated learning-based Edge-Cloud collaborative framework to incorporate pathological data diagnosis services at each Edge under both Labels-at-Edge and Labels-at-Server service scenarios. To adapt to the heterogeneous data settings from medical institutions, we introduce federated learning, allowing each Edge to upload its unsupervised and supervised parameters to the server model. Moreover, we introduce the Regrouping Averaging strategy in Fedsemgan to reduce gradient diversity. Comprehensive results demonstrate that our proposed approach significantly outperforms other contrast baselines.