In this study, a classification method is proposed based on a circular convolutional neural network (CNN) to deal with the issues of over-fitting and fine-grained classification of the malignant and highly deteriorated thyroid nodules. First, the Xception network and long short-term memory network (LSTM) are used as two non-interference parts. Next, the features of thyroid nodule samples are separately extracted to obtain two weight matrices, which are subsequently merged into a single weight matrix using the Merge algorithm. Further, the single weight matrix is imported into the CNN for feature extraction and pooling. Finally, the Softmax function of L2 regularization is used as a classifier to complete the training and testing of the circular CNN. Our experimental results denote that the accuracy of the fine-grained classification of malignant thyroid nodules is 87.00%, denoting a good feature-extraction capability.