Radiotherapy is recognized as one of the foremost treatment methods for cancer in clinical practice, to precisely deliver a prescribed dose to the planning target volume (PTV) while minimizing radiation exposure to surrounding organs at risk (OARs). Recently, deep learning methods have been widely applied in medical image processing. As a branch of medical image processing, dose prediction tasks also extensively utilize deep learning methods. However, current research has not proposed effective solutions to address the prediction discrepancy in difficult-to-predict regions. Therefore, we proposed a coarse-to-fine learning framework with uncertainty loss for dose prediction tasks. Specifically, this method divided the region to be predicted into easy-to-predict and difficult-to-predict regions. Then, by improving the prediction accuracy of difficult-to-predict regions, this method can improve the overall prediction accuracy. Meanwhile, due to the heterogeneity of data, the position of organs varies to a certain extent among different patients, with a greater difference in the position of the PTV compared to OARs, and considering that separating PTV and OARs for processing can achieve higher prediction accuracy. Therefore, both PTV and OARs are fed into the model separately to train the model. Experiments demonstrate that our method can predict more precisely dose maps.