Research On Medical Image Segmentation Of Multiple Types Based On Improved U-Net And A Novel Segmentation Process

Medical image plays an important role in the clinic.Obtaining organs or lesion areas from medical images quickly and accurately is the basis for targeted diagnosis.However,in reality,it is not easy for radiologists to quantify organs or lesion areas from medical images.It takes about 15 minutes to manually segment a medical image,which makes medical image segmentation a time-consuming and laborious work.Therefore,it is of great significance to invent automatic methods which can locate,segment and quantify organs or lesion areas.By using automatic medical image segmentation methods,the efficiency of medical image analysis can be effectively improved and the rate of clinical misdiagnosis can be effectively reduced.Recently,with the continuous development of deep learning,many medical image segmentation algorithms based on convolution neural network have been proposed.Among them,U-Net is one of the most famous medical image segmentation networks.However,the standard convolution layers adopted by U-Net limit its ability to capture rich features.In addition,the continuous maximum pooling operations in U-Net also lead to some loss of features in the segmentation process.For solving the above problems,a novel medical image segmentation model MHSU-Net based on improved U-Net is proposed.Firstly,in order to make the segmentation model capture rich semantic features,a novel multi-scale convolution module MCB is proposed.MCB adopts a wider and deeper network structure,which can extract various types of feature information,and can be applied to different types of medical image segmentation tasks.Then,a hybrid down-sampling module HDSB is proposed to reduce the feature loss in the segmentation process by replacing the maximum pooling layer in U-Net.Thirdly,this paper also proposes a novel context module CIF based on atrous convolution and adaptive receptive field mechanism in SKNet to obtain context information.Finally,this paper combines the CIF module with the skip connection structure in U-Net,and further proposes the skip connection+ structure,which not only improves the utilization of CIF,but also makes the context information obtained by the segmentation model more sufficient.In addition,during the process of segmenting some medical images,especially when segmenting tumor lesions,we usually encounter the problems of large scale difference and different area sizes of lesions.There are two reasons for the circumstance.First,patients are in different cancer stages,some in the early stage,some in the middle and late stage,resulting in different tumor sizes.Secondly,doctors may operate imaging equipment from different angles,resulting in certain differences in imaging results.Presently,improving the network structure(such as MHSU-Net)is usually adopted to alleviate this problem.However,the segmentation results still need to be improved.Therefore,this paper proposes a novel medical image segmentation process based on transfer learning.According to the idea of "first the whole,then the part",the first segmentation model is trained based on all the training images,and then this model is fine tuned on part of the training images by using the idea of transfer learning to deal with the segmentation tasks of large lesion region and small lesion region respectively,so as to achieve division and rule.In order to test the segmentation performance of MHSU-Net,this paper evaluates it on three different types of medical image datasets,including lung,cell contour,pancreas,and adopts evaluation metrics suitable for each segmentation task.The experimental results show that the proposed MCB,HDSB,CIF and skip connection+ can greatly improve the feature extraction ability of U-Net and effectively reduce the feature loss in the segmentation process.The segmentation results of MHSU-Net are indeed significantly better than the original U-Net.In order to test the effect of the proposed medical image segmentation process,this paper applies it to the skin lesion segmentation task,and selects appropriate hyper-parameter setting through comparative experiments.The experimental results show that the proposed medical image segmentation process is effective.