Research And Implementation Of Thyroid Ultrasound Image Classification Method Based On Medical Priori Knowledge

The incidence of thyroid disease has been increasing year by year,and it has become one of the diseases that plague people’s daily life.Early detection and treatment of thyroid disease is the key to cure it.Medical ultrasound imaging technology has become the first choice for thyroid disease detection and diagnosis due to its convenience,safety,reliability and other characteristics.However,due to its low contrast,high noise and high labeling cost,it has brought challenges to the detection and classification of lesions.In recent years,the rapid development of deep learning technology in the fields of image classification,detection and segmentation has brought new technical and theoretical support to the in-depth study of medical images.In this paper,thyroid ultrasound images were taken as the research object,based on deep learning methods.Based on the method of deep learning,the detection of thyroid nodules and the classification of benigh and malignant were further studied.First,the network structure and the implementation of benchmark model Faster R-CNN is analyzed in-depth,and Res At-Faster R-CNN is proposed.This model is improved on the basis of the original network: 1)Residual unit is embedded in the feature extraction network.The unit increases the width of the network and reduces the number of network training parameters.To some extent,the problem of gradient disappearance is alleviated.2)Squeeze-and-Excitation module is introduced,which can effectively extract contextual information.The block can increase the weight of the effective area of the image by learning through the network and it has the role of position correction;3)Using bilinear interpolation algorithm,the feature mapping process in the original model is changed,and the mismatch of ROI local regions of thyroid nodules caused by quantization and integration is improved.Then based on the characteristics of thyroid physiological and anatomical structure,the model is further improved by combining with medical priori knowledge:1)According to the characteristic that thyroid nodules are generally distributed near the center of the image,a priori constraint based on location is introduced.The screening rules of anchor box is re-formulated by supplementing different weights to the distance from the center of the anchor box to the center of the image and the score of the classification,in order to obtain more accurate candidate regions;2)For the characteristics of thyroid nodules with different sizes,the concept of multi-scale is introduced,and the Feature Pyramid Network is improved.This kind of structure can effectively improve the performance of the model on multi-scale nodules through cross-layer connection and fusion with high-level semantic information and low-level location information.Finally,we verified the rationality and effectiveness of the above improvement strategies from many aspects and angles through a large number of experiments on medical private dataset.The performance of the model has been significantly improved.Based on the research results of this subject,a thyroid ultrasound image classification system is designed and implemented,which can help doctors detect nodules and classify benign and malignant ultrasound images uploaded in real time in clinical diagnosis and treatment.