Ultrasound Diagnosis Model Of Benign And Malignant Thyroid Nodules Based On Multi-Task Deep Learning

It’s significant that use ultrasound images to assist doctors in diagnosing benign and malignant thyroid nodules for the early treatment of patients with thyroid cancer.Aiming at the problems of poor ultrasound image quality and few samples,difficulty in extracting deep features of nodules,low recognition rate of benign and malignant,and insufficient use of prior knowledge,a multi-task deep learning ultrasound-assisted diagnosis model of benign and malignant thyroid nodules was proposed.The main research content of the project and the phased results obtained are as follows.Firstly,carry out research to improve the quality of the data set.Preprocess the original ultrasound image,extract the region of interest in the thyroid ultrasound image,remove the artificial markers for less noise,and select the adaptive gamma correction with weighting distribution to increase the contrast,thereby improving the quality of the ultrasound image,and facilitating the expansion of more high-quality ultrasound samples.Secondly,carry out research on the construction of recognition models based on multi-task deep learning.It is proposed to build and train a multi-task deep convolutional neural network,which is divided into two steps: the first is to select and introduce Efficient Net,Xception,and Inception-v3 with outstanding classification performance,and train these convolutional network sensitive to nodules based on the transfer learning and fine-tuning strategy,concatenate into a multi-scale deep feature extraction module,so as to avoid the limitation of single scale and model structure on deep feature extraction ability.Transfer learning also improves the ability to capture differentiated features with fewer samples.the second is constructing multi-task learning model based on U-shaped network: first pre-train an improved segmentation task network with better performance,and then merge the feature output of the deep feature extraction module in the median layer,and jointly fine-tune the training classification capabilities of the shared layer and the fully connected layer to achieve nodule segmentation and classification,while sharing parameters to improve the generalization performance of the model.Thirdly,carry out research on the use of prior knowledge of thyroid to further improve the recognition effect of the model.A quantification method for assistant diagnostic features is proposed,and lesion masks based on automatic network segmentation is used to adaptively quantify the aspect ratio and circular compactness of nodules,and participate in classification as pathological features that affect the decision of benign and malignant nodules,so as to make full use of medical a priori Knowledge,further enhance the effect of nodule benign and malignant recognition.Finally,a more comprehensive comparison experiment and result analysis of the model were carried out.Comparative experiments show that the multi-task deep learning model is better than other same-target classification algorithms on the real ultrasound data set for nodule diagnosis.When combined with the multi-task assistant diagnosis method,its accuracy rate is 95.3%,and the sensitivity is 97.5%.The specificity is 91.0%,and the AUC is 0.989.At the same time,segmentation indexes such as Dice coefficient are also superior to the same type of segmentation network.This means that in the case of uneven distribution of benign and malignant samples,the ultrasound-assisted diagnosis method of thyroid nodules based on multi-task deep learning is effective and can assist doctors in diagnosing the nature of thyroid nodules faster and more accurately.