The fidelity of dose distribution prediction is paramount for radiotherapy planning. While existing deep learning-based methods have obtained noteworthy performance, most of them pursue the accurate prediction of global dose distribution but neglect local regions with sharp variations in dose, leading to inadvertent irradiation of healthy tissues. Thus, this paper proposes a dose stratification method to confront this challenge, refining neural network predictions of dose distribution in a hierarchical manner, where low-dose regions will not be overshadowed by high-dose regions in loss calculation. More specifically, the dose distribution is stratified into four subcomponents predicted individually, and the ultimate dose distribution emerges from the amalgamation of these subcomponents. Furthermore, a homogeneity index-based loss function is designed to augment the homogeneity of dose distribution, thereby mitigating collateral impact on healthy tissues. According to the experimental results on head and neck cancer cases in the OpenKBP dataset, the proposed method outperforms state-of-the-art methods for dose distribution prediction. Notably, the proposed method predicts dose distributions aligning more closely with clinically viable plans, enhancing the credibility and interpretability of artificial intelligence in the domain of radiotherapy planning.