Ultrasonic Thyroid Segmentation Algorithm Based On Deep Learning

Thyroid is the largest endocrine gland in the human body,many diseases related to it mostly involve the shape and size of the thyroid.Conventional ultrasound examination of thyroid tissue size is also the basis for the diagnosis of thyroid disease.Therefore,the segmentation and volume reconstruction of thyroid tissue is of great significance for clinical pathological diagnosis.However,for the task of analyzing pathology,manual annotation of ultrasound sequence image slice by slice has high labor cost and is greatly affected by human subjective factors,which leads to large error in image annotation.Therefore,it is necessary to design a computer-aided diagnosis system which can automatically segment ultrasonic thyroid tissue,which has important clinical value in clinical medical research and pathological diagnosis analysis.With the continuous application of deep learning in the field of medical image segmentation,combined with computer-aided diagnosis system to build a computer vision and depth of learning has become an important research direction.Compared with traditional segmentation algorithms,which usually need manual intervention,deep learning can automatically capture high-resolution semantic information to achieve image segmentation task,which can effectively alleviate the workload of manual annotation and solve the impact of annotation error.Therefore,this dissertation uses the deep convolution network to construct the ultrasonic thyroid segmentation model,combined with the problems existing in the ultrasonic thyroid image,proposes the following solutions for this dissertation:(1)Since the ultrasonic images are converted from acoustic signals,and different imaging methods of different machines may lead to significant differences in the generated ultrasonic images.If the data sources are different,it is easy to lead to great differences in network training and difficulty in fitting the training model parameters.Therefore,this dissertation introduces the knowledge of traditional image omics and image frequency domain analysis,fuses the enhanced texture feature and edge feature in the high and low frequency domain to enhance the contrast of the image and weaken the difference between the images.At the same time,according to the different characteristics of spatial and semantic information contained in the abstract features of convolution at different depths,the method of mixed up-sampling is adopted to refine the network feature restoration method,so as to enhance the acquisition of spatial location information and local semantic information during feature restoration so as to improve the segmentation accuracy of the network.(2)Due to the acquisition of serial images by ultrasound examination,the target scales in the thyroid section images are not evenly distributed,and the target objects in the images are significantly different in shape,size and position,which leads to the weak recognition ability of the network for multi-scale target objects and inaccurate segmentation.Therefore,this dissertation designs a cascading multi-scale dilated convolution pyramid module to enhance the network’s ability to capture target sizes of different sizes by integrating global semantic information under different sensory fields.At the same time,a local attention module combining channel attention and spatial edge attention is designed to enhance the attention of key semantic channels in high-dimensional features and regions of interest in low-dimensional features,so as to improve the extraction of key target features by the network.(3)Since there are tissues and organs around thyroid tissues with small differences and significant margins in the sequence images,thyroid characteristics in the small target area may be covered.Therefore,aiming at the non-significant target segmentation task of background feature interference,this dissertation designs a dynamic multi-scale dilated convolution module,fuses global semantic information and local detail information of different sizes under different sensing fields,extracts multi-scale target objects and weakens the coverage of background features.A feature fusion attention module is designed to guide the network to pay more attention to the target area and optimize the acquisition of key features between high and low dimensional features.The experimental results show that the proposed method compared with the same medical image segmentation network,the accuracy,recall and precision rates are greatly improved.