Novel Neural Networks For Automatic Thyroid Ultrasound Image

Thyroid lies in the middle of the neck of the human body,whose function is to secrete and synthesize thyroid hormone.Ultrasound imaging technology has become the main way to diagnose thyroid diseases due to its real-time,inexpensive,non-invasive and non-radioactive characteristics.The shape,volume,boundary and texture of the thyroid are important criteria for diagnosis.However,the inherent characteristics of ultrasonic image,such as speckle noise,artifact,attenuation,low contrast and signal loss,increase the difficulty of thyroid ultrasound image segmentation.Under the condition of different complexity of neural network models and different computing resources,this paper studies the neural network model suitable for thyroid ultrasound image segmentation.In the case of shortage of computing resources,the neural network model with low complexity should be chosen.In order to solve the performance bottleneck of low complexity neural network model.This paper proposes an automatic thyroid ultrasound image segmentation method based on adaptive width radial basis function neural network.This model improves the segmentation performance of low complexity neural network model and achieves the performance close to that of deep learning model.The whole segmentation process is divided into image pre-processing,feature selection and extraction,radial basis function neural network classifier and image post-processing.A feature select method is proposed to obtain discriminative and low correlated features from a amount of handcrafted features.These selected features improve the segmentation performance and reduce the complexity of the adaptive width radial basis function neural network classifier.The adaptive width radial basis function network model combines the global context and local characteristics of the feature space to enhance the data fitting ability of adaptive width radial basis function neural network and improve the segmentation accuracy of the segmentation model.In the case of sufficient computing resources,the neural network model with high complexity can be selected.Due to the inherent characteristics of ultrasonic images,even the deep neural network model cannot achieve excellent segmentation results in the fuzzy boundary region.This paper proposes a deep contour model based on local region image characteristics.The segmentation results based on the deep neural network model are applied as initial contours of active contour model to further fine-grained segmentation.Deep contour model combining the deep neural network model and active contour model overcome over-fitting and under-fitting of thyroid region boundary segmentation of deep neural networkmodel,and the defect that active contour model requires initial contours.The active contour model determines the evolution direction of the contour and makes the contour move to the target boundary based on the local region image characteristics of the contour points,so as to obtain more accurate segmentation results.In this paper,a amount of experiments are designed to verify the performance of thyroid ultrasound image segmentation method based on adaptive width radial basis function neural network model and the deep contour model based on local region image characteristics.The training speed is less than one second,and the test speed is up to five frames per second,and the segmentation results are competitive with that of the deep neural network models.The segmentation accuracy of the deep contour model is high,which has certain ability to segment the fuzzy boundary of thyroid gland,and the limitation of small data volume on the performance of deep neural network model is broken.However,the depth active contour model has high complexity,large number of parameters,great demand for computing resource and slow testing speed.This model is not suitable for the environment with shortage of computing resources.