| As the only vascular image that can be obtained without pouring into the human body,retinal blood vessels image can provide high clinical value for the evaluation of systemic diseases.It is very necessary to achieve efficient and accurate segmentation of retinal blood vessels.However,the color retinal images obtained often contain artifacts and noise due to the imaging equipment and external factors,which is hard for doctors to segment the blood vessel.Traditional blood vessels segmentation using unsupervised method makes the imperceptible blood-vessel segmentation subsequent and discontinuous.It is not conducive to doctors’ follow-up judgment.With the increasingly excellent performance of deep learning in the field of computer vision,automated segmentation of fundus blood vessels using deep learning is currently recognized as the most accurate and efficient segmentation method.In view of the current problems of less public fundus data sets and inaccurate segmentation of fundus blood vessels,this paper proposes the use of deep learning to achieve accurate segmentation of images.The main works of this paper are as follows:(1)By studying traditional segmentation algorithms such as threshold region growth k-means clustering,this paper verifies that the use of traditional algorithms to segment retinal blood vessels will lead to problems such as discontinuous and inaccurate cavity segmentation in vascular segmentation.In addition,this paper summarizes and analyzes the traditional segmentation algorithms.(2)Based on U-Net model which is widely used in semantic segmentation,this paper proposes a multi-scale feature extraction segmentation model called Mixed-Net,using convolution kernels of different receptive field extract global information and local information of feature maps in encoder part.The use of skip connection across stages in Mixed-Net can fully fuse shallow and deep semantic information.In addition,the skip connection can reduce the semantic defect in feature splicing and enhance the continuity and accuracy of fine vessels segmentation.Ablation and comparison experiments prove that mixed-Net has stronger feature extraction ability than U-Net and other segmentation methods.The accuracy rate and AUC value on DRIVE dataset are 0.9461 and 0.9623 respectively.(3)Due to the protection of patient privacy,there are few public retinal blood vessels data sets that can be obtained,which will lead to over-fitting in training.Therefore,this paper proposes a recovery algorithm of the retinal image based on pix2 pix,called Mix GAN.The generator of MixedNet can extract more features with binary segmentation labels as input.Mix GAN obtains realistic retinal images with label without manual segmentation,which can double the size of the retinal blood vessels data set.(4)This paper proposes a medical image segmentation system in the actual medical scene.Firstly,Mix GAN is used to generate medical images with labels to amplify the data set,and then the amplified data set is used to train the segmentation model.Using the amplified data set to train the mixed-net network,the accuracy of 0.9471 and 0.9630 AUC values can be obtained in the retinal blood vessel segmentation task,and the accuracy of 0.9814 can be obtained in the ISBI cell segmentation task.Both of the promotion of segmentation performance proves that the medical image segmentation system in this paper can effectively improve the accuracy of medical image segmentation.It provides a method to improve the generalization ability of segmentation model in medical scenarios with insufficient data. |
