Research On Medical Image Segmentation Algorithm Based On Deep Learning

Medical image segmentation,as a part of medical image analysis,is the basis of tasks such as tissue or lesion detection and recognition,and plays an important role in computer-aided diagnosis tasks and intelligent medicine.With the development of deep learning technology,medical image segmentation based on deep learning is the general trend.The construction of medical image segmentation algorithm with high segmentation accuracy and intelligent auxiliary diagnosis system with simple and efficient operation can greatly improve the efficiency of accurate diagnosis of doctors and narrow the gap of medical level in different regions,and at the same time provide better medical services for people's growing health needs.Based on fundus images and ultrasound images of medical images as the breakthrough point,this article explores the medical image segmentation algorithms,builds an easy-to-operate intelligent auxiliary diagnosis system,conducts a large number of experiments on the DRIVE,STARE and HC18 data sets,aiming at the problems of low segmentation accuracy caused by small amount of data,uneven image quality and small structure of segmentation target in data sets.The following research work is carried out:(1)A fundus image segmentation algorithm combining attention mechanism and conditional generation confrontation network is proposed to improve the segmentation accuracy of smaller retinal blood vessels.Firstly,an attention module is designed to focus on useful features and suppress irrelevant features in the high and low level feature maps,and it is applied to the high and low levels of the network to focus on blood vessel extraction,thus improving the precision of small blood vessel segmentation;Secondly,the attention generator and the residual module are constructed into a conditional generative adversarial network to make the generator and the discriminator antagonistic to each other and promote the improvement of their capabilities;Finally,the BN layer in the attention generator is removed to alleviate the impact of unstable batch statistics under small batch training.The test accuracy,sensitivity,specificity,and AUC indicators on the DRIVE data set are 95.59%,82.88%,97.45%,and 97.86%,respectively.All indicators are better than the current mainstream retinal vessel segmentation methods.(2)A medical image segmentation algorithm to alleviate the loss of useful information is proposed.Firstly,a multi-scale adaptive detail feature fusion module(MSADFF)was designed to adaptively fuse multi-scale features to fully capture details at different scales to alleviate the loss of detail information;Secondly,dual-path up-sampling modules(DPUS)are designed to compensate information from space and channels and restore image resolution,thus effectively obtaining better segmentation results;Finally,an improvement is made on the basis of U-NET,and the multi-scale feature map is adaptive feature fusion through MASDFF before jump connection,and DPUS is used as the new upsampling method.The test accuracy,sensitivity,specificity,AUC,and F1-Score indicators on the DRIVE data set are 95.67%,82.20%,97.63%,98.05%,and 82.87% respectively;the test accuracy,sensitivity,and sensitivity on the STARE data set Specificity,AUC,F1-Score indicators are 96.50%,82.39%,98.25%,98.65%,83.88%,respectively.All indicators are better than the mainstream retinal blood vessel segmentation algorithms compared.In addition,the Io U index on the HC18 fetal ultrasound image data set can reach 95.42%,which is 0.61% higher than the U-Net network.(3)Design and implementation of medical image aided diagnosis platform.On the basis of the proposed segmentation algorithm,from the perspective of practical application,build a medical image-assisted diagnosis platform.The entire platform integrates intelligent management of patient medical information and intelligent diagnosis of doctors,with rich functions,simple operation,strong interaction,and strong application value,and greatly improves the efficiency of doctors' diagnosis.