Research On Benign And Malignant Thyroid Nodules Based On Multimodal Ultrasound Images

Ultrasonography is one of the main imaging methods for diagnosing thyroid nodules.Automatic differentiation between benign and malignant modules in ultrasound images can greatly assist inexperienced clinicians in their diagnosis,has the great significance for early screening of thyroid cancer.The key of problem is the effective utilization of the features of ultrasound image.In this study,we propose a method that is based on the combination of conventional ultrasound and ultrasound elasticity images based on a convolutional neural network and introduces richer feature information for the classification of benign and malignant thyroid nodules,reducing and avoid unnecessary puncture biopsies.In this study,1156 thyroid ultrasound images were analyzed and collated in 233 patients.Capture single-frame ultrasound image from ultrasound device videos and labels region of interest for nodules.The boundary of the region of interest in ultrasound image is automatically extracted by color channel change and corrosion expansion operation,and the image of the region of interest in conventional ultrasound and ultrasound elasticity is extracted according to the coordinate symmetry of the image.Then,convolution network model is pre-trained on large-scale natural image datasets,and transfer parameters to ultrasonic image data domains to generate high-dimensional features by fine-tuning.The classification performance loss caused by the transformation of input image size is solved by introducing the spatial pyramid pooling and global average pooling structure.The deep features of conventional ultrasound and ultrasound elasticity are combined to form a hybrid feature space and classify on the hybrid feature space.Finally,a method of classification of thyroid nodules based on multimodal ultrasound images is proposed.The edge texture features of shallow convolution shared images,the abstract semantic features of high-level convolution are related to specific classification tasks,and the problem of insufficient data on ultrasound image can be solve by using the transfer learning.At the same time,ultrasound elasticity imaging can objectively quantify the hardness of the lesions of thyroid nodules,combined with the texture features of conventional ultrasound,the hybrid features can more fully describe the differences between different nodules.The experimental results show that the accuracy of this method is 0.9470,which is better than other single data source methods.This study starts from the different imaging principles of conventional ultrasound and ultrasound elasticity,the different feature distributions are combined and achieves state-ofthe-art results.As the computer-assisted diagnostic system,it can effectively assist radiologists in diagnosing thyroid nodule disease in clinical practice,and further improve the accuracy of clinical diagnosis.