Research On Thyroid Nodule Ultrasound Image Diagnosis Method Based On Attention Mechanism

Ultrasound examination is a non-invasive,low-cost,and real-time medical examination method that has become the standard method for characterizing nodular thyroid diseases and is widely used in medical diagnosis.The application of computer-aided diagnosis systems based on deep learning to intelligent diagnosis of thyroid ultrasound images can provide auxiliary information for diagnosis and significantly improve the initial screening efficiency of physicians.However,the current intelligent diagnosis method for thyroid nodules has not fully utilized the data of the Contrast-Enhance Ultrasound(CEUS)modality.The complexity of its spatial-temporal enhancement mode brings more information while increasing the difficulty of feature extraction and representation.Therefore,the study of a thyroid nodule ultrasound image diagnostic method based on attention mechanisms is of great significance for improving the diagnostic accuracy in the CEUS modality by integrating multi-modal ultrasound data,mapping the thyroid position information obtained from the grayscale ultrasound modality to the CEUS modality.This can improve the initial screening efficiency of thyroid nodular lesions and reduce medical costs.Based on the analysis of multi-modal thyroid ultrasound images,this article proposes a feature fusion method that considers the interference of the background in grayscale ultrasound images.To address the problems of information loss and low interaction efficiency during feature fusion in the feature pyramid network,a cross-scale attention interaction feature fusion method is proposed.This method applies self-attention weighting to information at the same level,collects distributed multi-level global context channel information to extract discriminative features,and uses content-aware sampling and spatial attention to re-weight and fuse adjacent-level features,effectively fusing information from different levels,enriching the semantic and pixel-level information of each layer’s features.To address the problem of overfitting caused by the limited scale of thyroid medical image data,a knowledge distillation-based method is proposed to enhance the localization model for thyroid ultrasound image regions of interest.Breast ultrasound data is introduced to design a t-distribution mask that focuses on the real glandular region.The method constructs a foreground-background distribution distillation loss function guided by channel-space attention and a global information distillation loss function based on the Gram matrix,which guides the student model to learn the feature extraction ability of the teacher model and strengthens the localization model.To address the difficulty in feature extraction from Contrast-Enhanced Ultrasound(CEUS)modality data,a thyroid nodule ultrasound image diagnosis method based on attention mechanism is proposed.By using the position information provided by grayscale mode to map the thyroid region in the CEUS modality data,the proposed method improves the 3D Swin Transformer architecture by decomposing it into a 3D convolutional network for aggregating local features and a Video Vision Transformer(VVi T)for global modeling.The proposed method extracts features from both temporal and spatial domains of CEUS data,achieving the diagnosis of thyroid nodules’ benign and malignant.The experiment was conducted using a joint dataset of thyroid ultrasound images,and the results showed that the proposed cross-scale feature interaction network based on attention mechanism combined with multiple detectors achieved good performance,with a 15.14% improvement in average precision(AP)compared to the basic feature pyramid network(FPN).The proposed ultrasound image region of interest(ROI)localization model enhancement method based on knowledge distillation achieved a6.35% improvement in localization accuracy compared to the AGD method,while reducing computational costs and speeding up processing speed,ultimately achieving an AP precision of 0.944.The proposed thyroid nodule ultrasound imaging diagnosis method based on attention mechanism showed better performance on the ultrasound contrast dataset compared to 3D convolutional networks and pure VVi T networks,with an accuracy of 84.8%for benign and malignant nodule diagnosis,improving the accuracy of thyroid ultrasound imaging for nodule-related diseases.