Convolutional Neural Networks And Transformers For Auxiliary Diagnosis Of Thyroid Ultrasound Images

Thyroid cancer is one of the top five diseases worldwide,and its incidence rate has been increasing in China in recent years.Generally,preoperative examination for thyroid cancer involves fine-needle aspiration,which may result in excessive puncture.Therefore,scholars in the medical field have proposed the use of ultrasound examination,which can effectively avoid the negative factors that need to be considered during fine-needle aspiration.However,the results of ultrasound examination rely on the doctor’s judgment,and there are issues with high noise and low contrast in ultrasound images,which increase the difficulty of the doctor’s judgment and lead to misdiagnosis.To address these issues,this article proposes the use of Convolutional neural networks and Transformers for auxiliary diagnosis of thyroid ultrasound images from the perspectives of classification and detection.This dataset comes from a collaboration with a hospital’s thyroid patient database.A total of 945 patients were selected based on their pathological reports.The dataset consists of DICOM format medical images,which were preprocessed and divided into four datasets: Thyroid-2D,Thyroid-3D,Thyroid-2D-T,and Thyroid-3D-T.The Thyroid-2D dataset contains 763 thyroid ultrasound images,including 366 images of benign thyroid nodules and 397 images of malignant thyroid nodules.The Thyroid-3D dataset contains a total of 110,993 ultrasound images,including 56,000 images of benign thyroid nodules and 54,993 images of malignant thyroid nodules.The Thyroid-2D-T dataset consists of 397 images,including 184 images of thyroid with lymph node metastasis and 213 images of thyroid without lymph node metastasis.The Thyroid-3DT dataset contains a total of 54,993 images,including 26,527 images of thyroid with lymph node metastasis and 28,446 images of thyroid without lymph node metastasis.For the classification task,this study uses the Res Net-50 and Vision Transformer network models.In the Res Net-50 network,an improved Inception module is added,which effectively increases the network model’s receptive field and reduces the number of parameters.The improved residual structure and loss function are also added,reducing the problem of vanishing gradients.The Vision Transformer model’s unique multi-head attention mechanism can combine learned different features and maintain the spatial information between input image blocks by partitioning the image and adding positional parameters to each image block.The performance of the two networks varies on different datasets.For small datasets,Res Net-50 performs better with an accuracy of up to 90.79%,while for large datasets,Vision Transformer performers better with a prediction accuracy of 91.73%.For the detection task,this study uses the YOLOv5 and Swin Transformer network models.In the YOLOv5 network model,squeeze and excitation attention modules and hybrid attention modules are introduced to enhance the network’s feature extraction ability.In the Swin Transformer network model,unique hierarchical operations and sliding window strategies are used,with hierarchical operations reducing network computation complexity and information loss,and sliding windows enabling information interaction between different windows,to some extent expanding the receptive field.Finally,the performance of the two networks is tested on datasets of different sizes.The improved YOLOv5+SE performs better on small datasets,with an average precision(m AP)of 80.3%,while the Swin transformer performs better on large datasets,with an m AP of 83.8%.The experimental results show that the Convolutional Neural Network and Transformer network proposed in this article have different performance on datasets of different sizes.For small datasets,the convolutional network performs better,while for large datasets,the Transformer network model performs better.