<p>Thyroid cancer is one of the most common types of cancer, pathological diagnosis based on Fine Needle Aspiration Cytology is clinically used as the standard for assessing thyroid cancer. However, the complex structure and large-scale data volume of thyroid pathology images pose challenges in terms of accuracy and efficiency for automatic diagnosis. To address this practical problem, this paper proposes a knowledge distillation method called Multi-Dimensional Knowledge Distillation, which involves feature-based distillation and response-based distillation.We employ a 12-layer Vision Transformer as the teacher model. Feature-based distillation integrates feature information from spatial, channel, and class token, while response-based distillation is achieved through alignment with targets. We integrate information from these diverse dimensions and compress the knowledge into a 3-layer Vision Transformer, which serves as the student model. The student model is trained and evaluated using a dataset containing 22,111 thyroid cytopathological patches. Ultimately, our student model attains a Top-1 classification accuracy of 94.87%. Compared with the teacher model, there is only a 0.55% gap in accuracy, while the computational complexity of the model has decreased by approximately a factor of four. In addition, our method is capable of substantially inheriting the generalization advantages of the teacher model. These results collectively demonstrate the effectiveness of Multi-Dimensional Knowledge Distillation in knowledge transfer.</p>