Research On Lightweight Aided Diagnosis Algorithm Of Thyroid Nodules Based On Ultrasound Images

Objective:Thyroid cancer is one of the most common cancers,and its initial symptom is thyroid nodule in the neck.In recent years,many related computer diagnosis algorithms have been proposed by domestic and foreign researchers to assist in the diagnosis of thyroid nodules.Although these algorithms have achieved high accuracy,the deep learning models used in the algorithms are complex,with plenty of parameters and computations.The large number of parameters can easily limit the running devices in terms of hardware,thus preventing it from working properly.The large number of computations leads to long inference time,which is not conducive to the real-time diagnosis.In addition to accuracy,device universality and inference speed are also important factors in determining whether the model can be practically applied in clinical practice.This paper carries out the research on aided diagnosis algorithm of thyroid nodules based on ultrasound images,constructing the lightweight models for thyroid nodule diagnosis,so as to help clinicians screen thyroid nodules and improve the accuracy and efficiency of diagnosis.Methods:Firstly,in order to train the models more adequately and improve their generalization ability,data augmentation is performed by horizontal and vertical shifting and random rotation in the range of 0-360degrees.Since the ultrasound images of thyroid nodules used in the experiments have low brightness and poor contrast,contrast limited adaptive histogram equalization algorithm is used to improve the contrast between the main part of the thyroid nodule and the surrounding background,highlighting the lesion area.Secondly,a lightweight residual network,EDSRes Net,is proposed based on Res Net-34.The residual blocks are simplified by depthwise separable convolutions,and the convolution structure is redesigned to retain more feature details and make the model more robust.Attention mechanism is introduced into the residual blocks to extract effective information from the feature maps in the channel dimension.Finally,a convolutional neural network based on dual-pooling compression attention mechanism,HETDCNet,is proposed.Take Dense Net-121 as the backbone to strengthen the reuse of image features.An attention mechanism module,DCAM,is proposed,which uses two kinds of pooling to compress the feature maps,and generates the weight vector through the fully connected layers and the activation function layers.DCAM can better focus on the texture information during weight extraction,thus improving the model’s ability to recognize thyroid nodules.Heterogeneous kernel-based convolutions are used to replace the standard convolutions of the model to reduce the parameters and computations.The dense block structure is improved to optimize the process of gradient descent and enhance the ability to learn features.Results:The accuracy,sensitivity and specificity of EDSRes Net in the ultrasound image dataset of thyroid nodules are 92.4%,94.5%and 91.7%,respectively,which are 0.9%,1.3%and 0.6%higher than those of Res Net-34,while the number of parameters and FLOPs are only 6.6%and 10%of those before the improvement.The accuracy,sensitivity and specificity of HETDCNet are 94.5%,96.3%and 92.5%,respectively,while the number of parameters and FLOPs are only 1.6×10~6 and 1.6×10~8.The experiment shows that HETDCNet is superior to all the contrast models in terms of accuracy and sensitivity,and is only inferior to VGG-16 in terms of specificity.Conclusion:In this paper,two deep learning models,EDSRes Net and HETDCNet,are proposed to automatically diagnose thyroid nodules from ultrasound images.EDSRes Net can identify the benign and malignant thyroid nodules with a small number of parameters and computations.Compared with EDSRes Net,HETDCNet has fewer computations and strengthens the reuse of image features,which further improves the classification performance.This study effectively solves the problems of large number of parameters and computations in the existing thyroid nodule aided diagnosis algorithms,and ensures the classification accuracy,which is of great practical significance for the current clinical diagnosis of thyroid nodules.