Abstract: Aiming at the problems of poor generalisation ability and difficulty in segmenting extreme malocclusions in deep learning-based malocclusion segmentation methods, a Semi-Supervised Learning-Based Dual-Stream Multi-Hierarchical Semantic Segmentation Method for Malocclusion Models is proposed. Firstly, for the large number of folds in the gingival region of malocclusion, the grid spectral clustering algorithm is improved to generate self-supervised signals through the cohesive grid hierarchical clustering; then, the neighbourhood features are extracted to enhance the key information of the malocclusion, and the local feature processing flow is designed through the multi-level residual blocks to refine the local boundaries of the malocclusion and the global feature extraction flow is formed through the linkage of the multi-scale bias attention to identify malocclusion semantic information; finally, the dual-stream features are fused together for the segmentation of the malocclusion. The experimental results show that with 40% of labelled data, the proposed method improves the accuracy by 9.29 percentage points compared with the representative semi-supervised method, and reduces the computational effort by 91.49%, which is close to the segmentation accuracy and efficiency of the better fully-supervised method, and it can better meet the demand for the development of the virtual orthodontic system in an intelligent way.