| With the progress of medical imaging technology,a large number of thyroid nodules can be detected by ultrasound images,but it is difficult for doctors to accurately judge benign and malignant nodules by images.Therefore,fine needle aspiration biopsy(FNA)is a common method for preoperative diagnosis of benign and malignant nodules.However,the diagnostic deviation is still large,at least 20%of thyroid nodules after FNA examination still can not be given a clear diagnostic recommendation.In recent years,with the development of molecular diagnosis,the detection of gene mutation in fine needle aspiration specimens can greatly improve the diagnosis results of thyroid nodules,but its sensitivity is poor,the workload is large,so it is not suitable for large-scale application.It is difficult to accurately judge benign and malignant nodules through ultrasound images,resulting in a large number of unnecessary biopsy and surgery,which brings great mental and economic pressure to patients.In recent years,with the excellent performance of deep convolutional neural network in natural image classification,deep learning has been proposed for medical image classification.At present,most of the computer-aided diagnosis systems of thyroid nodules focus on the diagnosis of benign and malignant images,but the simple diagnosis of benign and malignant images is lack of certain interpretability for patients and radiologists.Therefore,this paper proposes a model based on the fusion of multi-task convolutional neural network and traditional retrieval to achieve more accurate ultrasound image diagnosis of thyroid nodules.On the basis of implementation of the diagnosis of benign and malignant thyroid ultrasound images,the model also generates an ultrasound diagnosis report about this image through the fusion of multi-task neural network and traditional retrieval methods.The diagnosis report can make the diagnosis results of benign and malignant thyroid images more interpretable.Experiments show that our model can generate well-structured sentences and reduce the workload of doctors to a certain extent.The main work of this paper can be summarized as follows:1)This paper proposes an auxiliary diagnostic model for thyroid nodules.This model inputs a thyroid ultrasound image,analyzes the characteristics of the ultrasound image,outputs the classification results of the benign and malignant,echo and other aspects of the image,and assists radiologists to diagnose and classify thyroid nodules.2)A model based on multi-task cascaded convolutional neural network and traditional retrieval is proposed to generate a structurally fixed and accurate ultrasonic diagnosis report.In this paper,the classification results of multi-task neural network are used to form a structurally fixed ultrasonic diagnosis report.At the same time,the introduction of report database can make the structure of the generated report more diversified.3)This method can effectively classify thyroid nodule images and generate a descriptive report.Compared with the reports generated by other image descriptive methods,the report generated by this model performs better in all evaluation indexes. |
