Research And Application Of Deep Learning-based Assistive Diagnostic Method For Thyroid Nodule Ultrasound Images

The thyroid is an important endocrine organ in the human body and plays a vital role in maintaining normal metabolism.Thyroid nodules are a common thyroid disease.Ultrasound imaging is usually used for initial diagnosis of the thyroid,followed by grading using diagnostic criteria such as TI-RADS.The diagnostic process of ultrasound imaging heavily relies on the personal ability and diagnostic experience of doctors.Therefore,misdiagnosis or missed diagnosis may occur due to insufficient experience or fatigue of doctors,causing patients to miss the best treatment opportunity or endure the pain of puncture.This thesis carries out research on using deep learning to segment nodules in thyroid ultrasound images and further classify the segmented nodules using TI-RADS.A semantic segmentation model is proposed.The effects of various image classification models in identifying TI-RADS features of thyroid nodules are studied and compared,and a solution for classifying diagnostic features is proposed.Finally,an auxiliary diagnostic system for ultrasound images is designed and implemented to reduce the misdiagnosis rate of doctors during diagnosis.The main work content of this thesis is as follows:1.A U-Net semantic segmentation model improved by an external attention mechanism is proposed.The ability of the U-Net model to obtain global information is enhanced by adding an external attention mechanism.After experimental comparison,it was found that the proposed model outperforms classic semantic segmentation models such as FCN,PSPNet and U-Net in evaluation indicators such as Io U and Dice.Compared with the original U-Net model,it has increased by 1 percentage point in Io U and 2.7percentage points in pixel accuracy,and the effect of the model is less affected by lesion size.2.The study investigated an integrated learning method for TI-RADS classification of thyroid nodules in ultrasound images.This thesis compared the classification performance of 9 models on diagnostic criteria and ultimately selected the 4 best-performing models.By combining these four models through an ensemble learning division method,a superior classification model was obtained.The results showed that this classification model can provide the most advantageous identification effect in this thesis.3.A thyroid nodule ultrasound imaging auxiliary diagnosis system was designed and implemented for the semantic segmentation model and classification model proposed in this study.Users can submit ultrasound images to the system for lesion segmentation and classification to assist doctors in diagnosis and reduce errors and omissions during diagnosis.This thesis investigated the actual needs of imaging specialists during diagnosis and conducted research on thyroid ultrasound image segmentation and classification.Using deep learning,we achieved automatic segmentation of nodules in ultrasound images and classification according to TI-RADS diagnostic criteria and developed an auxiliary diagnostic system.The results show that this research has certain application value in assisting doctors in diagnosing thyroid nodules.