Research On Medical Image Classification And Generation Based On Capsule Network

Papillary thyroid carcinoma is the sub-category with the largest proportion of thyroid cancer,and it can occur at any age.In recent years,the number of people with disease in China has increased year by year.Therefore,it is particularly important to improve the accuracy of thyroid papillary cancer diagnosis.At present,medical imaging is a very important technical means for medical examination and diagnosis of thyroid cancer.It can assist physicians in diagnosing diseases in two aspects: one is to obtain medical images with medical equipment and directly understand and judge;The doctor interprets the medical image and checks for abnormalities.If there are abnormalities,the location,size,and degree of calcification of the lesion need to be quantified,and the classification and grade of the disease must be determined.Ultrasound images for diagnosing papillary thyroid carcinoma have low resolution and poor sharpness,and because of the complicated tissue structure around the thyroid gland and many interference factors,it is difficult to classify the lesion features of thyroid cancer ultrasound images.With the development of artificial intelligence technology in recent years,convolutional neural networks in deep learning technology have been applied to medical imaging.However,the accuracy of models based on convolutional neural networks to identify medical images is generally not high.In addition to the factors that are difficult to identify in the medical image itself,there are also structural defects in the model itself,which are prone to overfitting or underfitting,resulting in insufficient learning ability.The lack of existing data sets makes the use of deep learning technology to identify papillary thyroid cancer with low accuracy.The lack of data also hinders the development of artificial intelligence technology in the field of medical imaging.This paper uses the capsule network in deep learning technology as the basic network to solve the classification and generation of papillary thyroid carcinoma.This paper proposes the following solutions to the two problems of classification and generation:Aiming at the classification problem,this paper proposes a new network model based on the original capsule network and named it ResCaps network.The ResCaps network uses the residual module to make the capsule network model input more advanced features,and strengthens the abstract expression of the model.It is more conducive to the ResCaps network model to solve the problem of overlapping ultrasound image tissues.By performing experiments against the CNNCaps network,ResCaps the network can improve the accuracy of ultrasound image classification of papillary thyroid carcinoma.Aiming at the generation problem,this paper integrates Batchnorm and Dropsort methods between the last three fully-connected layers of the capsule network,uses the new Huber loss function,and names the network BDCaps.Batchnorm and Dropsort methods have the ability to effectively prevent overfitting,and can achieve the effect of improving the accuracy of the final generated image after connecting with the fully connected layer.The generated medical images are added to the original data,which enriches the diversity of the data set and enhances the robustness of the BDCaps network model.Experiments show that the two new networks,ResCaps and BDCaps,have improved performance compared to the original capsules,and have certain promotion effects on the classification and generation of papillary thyroid cancer.