Research And Application Of U-shaped Structural Network Model In Thyroid Nodule

Thyroid nodules are a common endocrine disorder that can affect the normal function of the thyroid gland and endanger human health,with the risk of cancer in malignant nodules.Ultrasound imaging is widely used in the field of thyroid disease detection.It is an important task to identify and segment the diseased nodal tissue in ultrasound images to assist in the subsequent diagnosis and treatment.However,due to technical limitations,there are noises and spots in thyroid ultrasound images,and the thyroid nodule itself is closely adhered to the surrounding healthy tissue,which makes the boundary of the thyroid nodule tissue area blurred and unclear,which increases the work of identification and segmentation difficulty.In addition,most ultrasound images have more than one nodule of different sizes,and the existing segmentation methods suffer from poor segmentation and even problems of missed cuts.In response to the above problems,this thesis constructs two network models for thyroid nodule recognition and segmentation based on the U-shaped structure,and the main research content are as follows:(1)To address the problem of adhesion of thyroid nodule tissue to other surrounding tissues,resulting in blurred boundaries of the nodule region,the network model which named DCA-UNet++ was constructed for thyroid nodule recognition segmentation.The model is based on the U-shaped structure U-Net++.Firstly,the dilated convolution module is used to expand the perceptual field of the recognition segmentation model in the encoding path phase,keeping the input feature mapping size and resolution constant and enhancing the ability to obtain broad contextual information.Secondly,in the feature extraction stage,DCA-UNet++ add an efficient channel attention module,dynamically adjust the feature weight of each channel,reducing the interference of irrelevant information in non-nodular areas in thyroid ultrasound images,and highlighting important and key feature information in the images.In addition,the design uses a hybrid double loss function to ensure the convergence performance of the model.Finally,several evaluation metrics such as recall rate and average interaction ratio were used to verify the accuracy of the DCAUNet++ network model,in which the recall rate reached 97.17% and the average interaction ratio reached 75.22%.(2)The network model MCL-UNet was constructed for thyroid nodule identification and segmentation in response to the problem of variable size and size of thyroid nodule tissue,which leads to missing nodule region segmentation.The model is based on the U-shaped structure U-Net.Firstly,an innovatively constructed multiscale convolution module is introduced in the encoding path stage to capture feature information of different sensory fields with unequal size convolution operations as the entry point,taking into account the features of both shallow and deep stages to enrich diversified information extraction.Secondly,in the decoding path stage,a long-term short-term memory module is introduced to capture the spatial correlation between each pixel in the thyroid ultrasound image,taking into account the feature information extraction of different distances,and making full use of feature mapping information.Finally,several evaluation metrics such as interaction ratio and Dice similarity coefficient were used to verify the innovativeness and accuracy of the MCL-UNet network model for recognition segmentation,in which the accuracy reached 97.27%and DSC reached 80.72%.