Abstract: The Chinese-thyroid imaging reporting and data system (C-TIRADS) provides a guideline for determining the risk factor of malignant thyroid nodules. In order to solve the problem of low detection accuracy of existing methods, a self-attention and self-distillation based malignant risk stratification detection method for thyroid ultrasound images in accordance with the C-TIRADS standard was proposed. First, by fusing the convolutional neural network (CNN) and the self-attention model ViT, the feature extraction ability of the model in the malignant risk stratification detection of thyroid nodules was improved; then the fusion of ViT and CNN was strengthened, and the spatial and channel bottleneck structure. Second, the interaction module of the spatial and channel was proposed, so as to achieve the feature fusion in the cross-dimension, cross-window, and cross-scalar scales. Third, the improved graph convolutions network was introduced to deal with the interconnections between the pathological labels, and the output feature vector was corrected to allow the model to learn the correlations between pathological labels. Finally, we proposed to train the model by semi-supervised learning using improved noisy student self-distillation, which solved the problems of the imbalance of samples among categories in the thyroid nodule dataset and the overfitting of small datasets. The ablation and distillation experiments were conducted on real datasets and compared with seven different methods. The results showed that the proposed method was effective, and the mAP of the method reached 93.5% with an IoU threshold of 0.5, and 90.5% for small nodules, which was a big improvement compared with the traditional CNN.