Research On C-TIRADS-Based Ultrasound Image Detection Of Thyroid Nodules With Deep Learning Model For Risk Stratification

Accurate assessment of the extent of infiltration in the tissues and organs around the thyroid gland in the neck is of great significance for the selection of treatment options and prognosis of patients.The Chinese Ultrasound Thyroid Imaging Reporting and Data System(Chinese-TIRADS,C-TIRADS)has the characteristics of simple classification and easy implementation,which is helpful for clinicians to diagnose thyroid nodules,especially for risk stratification of thyroid malignancies.In view of the lack of a computer-aided diagnosis system for ultrasound images of thyroid nodules based on C-TIRADS,the thesis is devoted to investigate the key technologies of this system and proposes a deep learning-based ultrasound image detection and risk stratification model for thyroid nodules.The main work and milestones are as follows.(1)Construction and pre-processing of thyroid nodule ultrasound image dataset.In response to the lack of C-TIRADS-based ultrasound image dataset of thyroid nodules,a series of dataset construction work was carried out with the assistance of specialized physicians,including C-TIRADS-based ultrasound image acquisition of thyroid nodules,preliminary preprocessing and annotation.Due to the problem of artificial markers in ultrasound images of thyroid nodules,an improved Scharr algorithm and image binarization were proposed to detect artificially marked edge points,followed by IDBP(Iterative Denoising and Backward Projections)algorithm for image restoration,and a restricted contrast adaptive histogram equalization(Contrast Limit Adaptive Histogram Equalization,CLAHE)algorithm and elastic transformation algorithm are used for data enhancement to construct a high-quality dataset that meets the model training needs.(2)Construction of ultrasound image detection and risk stratification model for thyroid nodules.A deep learning-based ultrasound image detection and hazard stratification model of thyroid nodules is studied for the detection and hazard stratification identification needs of thyroid nodules.Firstly,based on the analysis of the characteristics of several advanced target detection models,the DETR(Detection Transformer)model with ideal detection effect is selected as the benchmark model through comparison experiments of designed models.Secondly,the improvement of the DETR model was carried out to address the problems in the detection and risk stratification of thyroid nodules,such as the lack of satisfactory classification accuracy and the difficulty in detecting small thyroid nodules.The first is to study the method of obtaining multi-scale features of thyroid nodule ultrasound images,and propose the use of TRes Net-L as a feature extraction network and the introduction of multi-scale feature information and group convolution,aiming to improve the feature extraction capability of the detection model;the second is to design a parallel decoder structure for multi-label thyroid nodule ultrasound image classification needs to enhance the learning of correlation between pathological feature class labels,aiming to improve the multi-label classification of the detection model.aiming to improve the multi-label classification accuracy of the detection model;third,improve the loss function of the detection model,and propose a linear combination of Smooth L1-Loss and CIo U Loss as the bounding box loss function of the model and Asymmetric Loss as the multi-label classification loss function of the model,aiming to further improve the detection accuracy of the detection model for small thyroid nodules.Finally,the improved DETR model was experimented and analyzed on the ultrasound image dataset of thyroid nodules.The experimental results demonstrated that the enhanced DETR model achieved AP scores of 92.4% and 81.6% at Io U thresholds of 0.5 and0.75,respectively.These results confirm the model’s improved ability to accurately detect the location and risk level of nodules in C-TIRADS-based thyroid nodule ultrasound images.This provides a more precise and reliable diagnostic tool for clinicians assessing the risk level of thyroid nodules.