Automatic Segmentation And Classification Of Thyroid Nodules In Ultrasonography Based On Hierarchical Deep Learning

The proportion of thyroid cancer has been increasing in recent years.As a highincidence endocrine system disease,thyroid cancer has become one of the top ten cancers that endanger human health.Early diagnosis and recognition of the nature of thyroid nodules is conducive to the promotion of follow-up treatment.Ultrasonography has become the preferred method for clinical detection of thyroid cancer due to its advantages of safety,non-invasive and high diagnostic rate.The computer-aided diagnosis of thyroid ultrasound images is also helpful to break through the limitations of artificial experience and help doctors to analyze the lesions.However,due to the low imaging quality of ultrasound images,the serious speckle noise,and the complex and diverse ultrasonic manifestations of nodules,the classification of thyroid nodules is extremely challenging.In view of the above problems,this thesis proposes a hierarchical deep learning model based on mask guidance,which can be used to realize the automatic extraction of ultrasonic thyroid nodule mask and the automatic classification of benign and malignant nodules.Firstly,through an improved Mask RCNN model,thyroid nodule Mask was automatically extracted from ultrasonic images as the region of interest(ROI)for subsequent classification,and a series of related lowdimensional features of images were extracted on this ROI.Then,a deep network based on residual attention is designed to realize the deep feature extraction of ultrasonic nodule ROI,and the dimensional alignment technology was used to fuse the deep features with the low-dimensional features to form a mixed feature space.Finally,a convolutional neural network based on Attention Drop was used to realize the automatic classification of benign and malignant thyroid nodules in the mixed feature space.The comparative experiments on the public data set DDTI(Digital Database Thyroid Image)show that the IOU segmentation result is 93.75%,the average classification accuracy is 93.01%,the AUC value is 91.49%.Compared with the comparative methods,the proposed method in this thesis not only got a better classification results,and the hierarchical deep learning network can also improve the classification performance,which has certain clinical application value.