Research On Medical Image Segmentation Technology Based On Deep Learning

With the development of medical image technology,doctors began to use medical image data to complete the segmentation,3D reconstruction,classification and recognition of human tissues and lesions,and gradually formed a set of auxiliary diagnosis system,which greatly improved the accuracy and reliability of medical diagnosis.Ultrasound,CT and MRI are the most common imaging technologies in the field of medical imaging technology.Using these imaging techniques,doctors can quickly judge the patient’s condition and give the corresponding treatment measures and methods,which can provide fast and effective help for the treatment of the patient’s condition.Different from CT and MRI,ultrasound technology has the advantages of simple imaging technology,fast imaging speed,low data delay,good penetrability and real-time imaging,and it is widely used in thyroid,prenatal diagnosis and other medical examinations because of its small radiation and low price of ultrasound examination.Therefore,segmentation of ultrasound image,as a key step in the auxiliary diagnosis system,plays a very important role.At the same time,with the improvement of computing power and the increase of data,deep learning has made significant progress in the field of medical image segmentation.The automatic learning of local features and high-level abstract features through multi-layer network is better than manual extraction and prediction.Therefore,this paper will combine the advantages of ultrasonic image and deep learning to study the ultrasonic image segmentation technology based on deep learning.As follows:1.Design ultrasonic image preprocessing method.Data preparation: developed the image acquisition scheme of control variables,and completed the label of original ultrasound image under the guidance of doctors.Data processing: modern digital image processing technologies such as guided filtering,histogram equalization,Sobel edge enhancement are used to denoise the ultrasonic image,improve the contrast,binarize the label and divide the data set.Data enhancement: the online and offline methods are combined to expand the training set and verification set,reduce the memory pressure and enhance the generalization effect of the model.2.Design an ultrasonic image segmentation algorithm that can accurately segment medical tissues or lesions based on u-net network structure.Including the following three innovations:1)Multi-input densely dilated convolutional encoder module.The input data is scaled into four groups and concatenate with each subsampled layer.The dense convolution network is integrated,and the dilated convolution method is used to expand the receptive field and capture multi-scale information.This module alleviates the problem of gradient vanishing and enhances feature transfer and feature reuse.2)Pyramid attention center module.The attention mechanism and spatial pyramid pooling are combined to extract accurate and dense pixel level features and expand the range of receptive field through pyramid structure.This module solves the problem of different object scales and bad segmentation effect of small objects.3)Deep supervision channel attention decoder module.The feature extracted from the encoder stage is used as the gating information of the decoder stage,and the salient feature is transferred through the skip connection.This module can suppress the response of the region independent of the segmented target,and help speed up the convergence.The technical method of this paper integrates multiple models,such as multi-scale framework,densely dilated convolution network,attention gate,spatial pyramid pooling,new loss function,etc.,which is helpful to realize feature reuse,suppress the response of irrelevant area,improve the performance of small region of interest,solve the problems of small sample,low pixel,fuzzy boundary,large difference of ultrasonic image,and obtain the optimal segmentation.