Research On Intelligent Risk Assessment Technology Of Thyroid Nodule Tumor Based On Ultrasound Image

Thyroid nodules are lumps in the thyroid that are at risk of becoming malignant.Medical ultrasound is the most effective non-invasive testing tool for evaluating thyroid nodules.In clinical diagnosis,the Thyroid Imaging Reporting and Data System(TI-RADS)is the most commonly used risk assessment system for nodular malignancies.However,the use of TI-RADS is more complicated,and doctors need to have rich experience in diagnosis and treatment to make an accurate evaluation quickly and accordingly.Therefore,based on the ultrasound imaging and TI-RADS system,this dissertation studies the intelligent risk assessment technology of thyroid nodule tumors to ensure the accuracy of screening and improve the efficiency of inspection.The main research contents and innovations of this dissertation can be summarized in the following aspects.(1)Accurate detection of thyroid nodules in ultrasound images is the primary issue of intelligent evaluation.In view of the frequent misdetection phenomenon in nodule detection by existing models,this dissertation proposes a feature-enhanced dual-branch nodule based on multi-tasking The model is tested,and the effectiveness of the method is verified on the dataset.Nodule detection requires searching for suspected nodule regions from ultrasound images and obtaining their spatial locations.However,the existing target detection models frequently occur false detections in nodule detection,mainly because these models cannot stably obtain high-order implicit features of nodules from ultrasound images.Quality matters.This dissertation firstly analyzes the detection principle of Faster R-CNN in detail.Then,based on Faster R-CNN,this dissertation proposes a multi-task-based feature-enhanced dual-branch nodule detection network.The design comprehensively considers the feature similarity in different visual tasks,adds a segmentation task branch independent of the detection task in Faster R-CNN,and realizes the information interaction between the two tasks through the attention mechanism to achieve the enhancement of implicit features.Such a design can suppress the influence of the background in the convolutional features and effectively improve the accuracy of nodule detection.The experimental results show that the average accuracy of the proposed nodule detection method in the two datasets is 1.8%and 6.0%higher than that of the Faster R-CNN method,respectively.And the experimental results also verify the effectiveness of the feature enhancement method through multi-task.(2)Aiming at the problem that the model cannot be trained due to the lack of pixel-level mask annotations,this dissertation proposes a weakly supervised nodule segmentation method combined with nodule detection to predict nodule masks,and it is better than a variety of open-source schemes.Considering the inefficiency of mask annotation,obtaining large-scale thyroid ultrasound images with pixel-level masks for model training is difficult in practical scenarios.With the idea of seeding,dilation,and constraint joint loss function supervised training,this dissertation proposes a weakly supervised nodule segmentation training model for joint nodule detection.First,the seed loss function is designed by using the bounding box classification task in the detection network,which solves the problem of feature disappearance caused by the RoI pooling layer when the class activation map is generated by jumping.Second,when designing the dilation loss function,this dissertation proposes an algorithm of a random local average pooling layer to satisfy supervised training in the absence of negative examples.As for the constraint loss function,this chapter uses the ground-truth bounding box and the generalized intersection ratio loss function to constrain the bounding box of the prediction mask to limit the infinite expansion and discontinuity of the nodule region.By training with a joint loss function,the segmentation model can perform more refined mask predictions.In addition,combined with the common practice of weakly supervised training,the mask predicted by the segmentation network is used as a "fake" ground truth mask to iteratively train the feature-enhanced dual-branch nodule detection model,and finally,the mask prediction is refined and the nodule detection performance is improved..The experimental results show that the weakly supervised nodule segmentation method proposed in this dissertation is better than a variety of open-source schemes.Although the feature-enhanced dual-branch nodule detection model trained by using the "pseudo" ground truth mask cannot achieve the indicators under the fully supervised condition,It still outperforms the nodule detection performance of other models under full supervision.(3)In order to conform to the clinical diagnosis habits of doctors,this dissertation proposes a multi-dimensional intelligent malignant tumor risk assessment model for thyroid nodules based on TI-RADS.Compared with the direct classification of benign and malignant,the accuracy of this method is more obvious.improvement.TI-RADS selected 5 ultrasound features closely related to benign and malignant nodules to obtain the TI-RADS score of the nodule and selected "benign","not suspicious","mildly suspicious" and "moderately suspicious" according to the score."or " highly suspicious" to describe the risk of malignancy for the nodule,where the higher the value,the greater the risk of malignancy.Although TI-RADS is the most commonly used evaluation standard,the relevant analysis at this stage is still mainly focused on the direct classification of benign and malignant thyroid nodules,which cannot truly conform to the clinical diagnosis habits of doctors.This dissertation firstly designs an image feature cropping model that is more conducive to nodule classification according to the actual task needs.Then,on this basis,a multi-dimensional intelligent evaluation model was proposed to synthesize the classification results of different ultrasound features of nodules in TI-RADS and make a reasonable evaluation of nodules.In this model,HyFCNet will be trained as multiple independent classifiers to complete multiple sets of classification tasks while sharing the convolutional neural network feature results.The experimental results show that the average accuracy of the multi-dimensional intelligent malignant tumor risk evaluation model in this dissertation is higher than other methods on the five TI-RADS features,and the accuracy index of direct multi-class evaluation is improved from 79.10%to 88.45%,and then analyzed the correlation of different ultrasound features through the acquired implicit features.To sum up,this dissertation explores the problem of automatic detection of thyroid nodules in images based on thyroid ultrasound image data,and performs joint training with nodule segmentation task to optimize the effect of nodule detection;a weakly supervised learning segmentation for joint nodule detection is designed.The algorithm realizes the mask prediction of nodules and trains the detection model under the condition of weak labeling;realizes the multidimensional evaluation model of malignant risk of thyroid nodules based on TI-RADS.The research in this dissertation covers several key technical points in intelligent risk assessment of thyroid nodules,which will help to improve the reliability and availability of computer-aided diagnosis systems in this field.