Microscopic hyperspectral imaging can simultaneously capture spatial structural information and continuous spectral signatures of biological tissues, thereby providing a novel means for disease detection and analysis at the cellular level. By integrating microscopic morphology with spectral reflectance characteristics, this technique makes it possible to identify subtle pathological changes that are difficult to observe using conventional imaging methods. However, hyperspectral images are typically characterized by extremely high dimensionality, strong inter-band redundancy, and substantial data complexity. In addition, pixel-level annotation of microscopic hyperspectral images is highly labor-intensive and time-consuming, which greatly restricts the direct application of fully supervised segmentation methods, especially in plant disease studies where expert labeling is costly and limited.

To address these challenges, this study proposes a dual-model semi-supervised segmentation framework with dynamic high-confidence pseudo-labeling for microscopic hyperspectral images of wheat leaves infected by stripe rust. The proposed framework is designed to reduce the dependence on large-scale manually annotated datasets while improving segmentation accuracy and robustness. Specifically, the method employs a convolutional neural network (CNN) and a Vision Transformer (ViT) as two complementary teacher-student branches. The CNN branch is effective in capturing local texture details and fine-grained spatial patterns, whereas the ViT branch is more capable of modeling global contextual dependencies and long-range semantic relationships. Through a bidirectional cross pseudo-label supervision strategy, the two models iteratively exchange reliable predictions and jointly exploit their complementary advantages, enabling the framework to learn more discriminative representations from limited labeled samples and abundant unlabeled data.

Furthermore, a dynamic threshold update mechanism based on validation-set performance is introduced to adaptively adjust the pseudo-label selection threshold according to the evolving capability of the models during training. Unlike fixed-threshold strategies, the proposed mechanism allows the pseudo-label filtering criterion to change with model performance, which helps maintain a balance between pseudo-label quantity and quality throughout the learning process. At the same time, a high-confidence filtering strategy is incorporated to retain only those pixels with sufficiently reliable teacher predictions for student model optimization. This design effectively suppresses the accumulation of noisy pseudo-labels and reduces error propagation during iterative training, thereby enhancing both segmentation stability and generalization ability.

Experimental results demonstrate that the proposed method can significantly improve the segmentation performance of microscopic hyperspectral images of wheat stripe rust. After semi-supervised optimization, the U-Net model achieved an mDice of 0.8290 and an mIoU of 0.7079, while the ViT-based model achieved an mDice of 0.8209 and an mIoU of 0.6962. These results indicate that the proposed framework effectively enhances segmentation precision under limited annotation conditions. Overall, this study provides a feasible and effective solution for reducing manual labeling dependence in microscopic hyperspectral image segmentation. By integrating complementary model architectures with dynamic pseudo-label optimization, the proposed framework effectively alleviates pseudo-label noise and improves segmentation accuracy and stability. The method can provide technical support for wheat tissue structure analysis and host-pathogen interaction phenotyping research, and it also offers a useful reference for semi-supervised segmentation of other microscopic hyperspectral imaging tasks in agricultural and biological applications.