| Thyroid nodules are more common in the clinic,and their incidence is also increasing with the accelerated pace of life and poor living habits.Ultrasound diagnosis is a common way to diagnose thyroid nodules.In recent years,some machine learning methods have been used to study the benign and malignant classification of thyroid nodules ultrasound images,but the accuracy and robustness of the model need to be further improved.This paper proposes a classification method based on capsule network for the benign and malignant thyroid nodules,which is mainly divided into two parts: thyroid nodule localization and classification.In the study of thyroid nodule localization,this paper is divided into three parts: image preprocessing,artificial marker recognition and boundary adjustment: image preprocessing removes the content unrelated to the study and preliminarily determines the possible position of the marker;artificial marker recognition is used for judge the found marker and accurately identify the specific position of the marker;boundary adjustment estimates the not found marker and gives a rectangular frame for defining the nodule.Thyroid nodule positioning accurately locates the nodule and removes irrelevant information from the image for subsequent studies.In the study of classification of ultrasound images,this paper applies the newly proposed capsule network model to the classification of thyroid nodules.In this paper,the influence of feature vector layer on capsule network is studied.The distribution of feature vector is discussed.The method of dropout is introduced into the capsule network,and compared with convolutional neural network in the case of image translation and image rotation.Experiments show that the improvement of the capsule network proposed in this paper effectively improves the accuracy of the model and the adaptability to image transformation.It is feasible to apply the capsule network to the classification of ultrasound images.In order to illustrate the effectiveness of the improved capsule network,in addition to experiments on ultrasound images,related experiments were carried out on the MNIST dataset and the Fashion-MNIST dataset.The results of these experiments also show that the method of this paper has certain adaptability and effectiveness. |
