Image Segmentation Of Thyroid Nodules Based On Deep Learning

Thyroid nodule is a common disease of thyroid system,and malignant nodules indicates the coming of the thyroid cancer.For malignant nodules,most of them can be cured after timely diagnosis and treatment.Ultrasonography is the most commonly used diagnostic method for thyroid nodules.In order to solve the problems of time-consuming diagnostic film reading by doctors and prone to false detection and missed detection due to different personal experiences,computer-assisted diagnosis technology based on ultrasound image is of great significance in clinical practice.The aim of this paper is to achieve an end-to-end,efficient and accurate ultrasound image segmentation method of thyroid nodules.Traditional nodule segmentation methods are mainly based on image processing technology and active contour model.These methods fail to acquire knowledge of the professional characteristics of thyroid nodules and require artificial pre-targeting in practical applications.In this paper,a convolutional neural network model is proposed to achieve end-to-end accurate segmentation of thyroid nodules.Firstly,in order to train and validate the performance of model,we constructed the thyroid nodule ultrasound image data set and completed the dataset annotation under the guidance of professional medical personnel.To solve the problem that medical data are generally small in scale,we propose an image enhancement method based on elastic transformation and affine transformation for expanding the model training data set.Through comparative experiments,it is proved that this data enhancement method can significantly expand the dataset.Secondly,aiming at the convolutional neural network model used for nodule segmentation,we designed and constructed several control experiments of model structure.Through the experimental results,DenseNet is determined as the backbone of network,and U-net++,which is based on U-net,changes the single skip-connection into the fusion of multi-size feature maps,is used as the model’s structure of thyroid nodule segmentation.Inspired by multiple-model fusion method and in consideration of the low semantics of thyroid nodule images,we proposed a test time augmentation method,which can significantly improve the final segmentation result without training multiple independent models.Finally,considering the particularity of medical problems,we visualized the model from two perspectives,one is through the convolution layer output of the actual picture,the other is the maximum response of the convolution kernel.In this way,we enhanced the interpretability of the model during nodule segmentation.We also deploy the model to an embedded platform to achieve convenient and offline segmentation to ensure the security and privacy of medical data.