Research On Some Key Problems Of Medical Image Segmentation Based On U-Net

Medical images can provide disease details to doctors,which is sufficient for the diagnosis,treatment and prognosis analysis.Medical images can be segmented into several disjoint regions basing on some similar features,.The results of medical image segmentation play an important role in medical image visualization,lesion area recog-nition and surgical planning.In recent years,segmentation methods basing on deep learning have strong application potential in the field of medical image segmentation.At present,UNet is the most widely used framework in the field of medical image seg-mentation.However,it is difficult to get accurate segmentation results because of the category imbalance and noise interference in medical images.Moreover,how to im-prove the utilization of multimodal medical images is also a problem worth discussing.Therefore,some key problems in medical image segmentation have been deeply studied combination with the U-Net framework in this dissertation.Firstly,dual encoder-decoder network basing on multiscale blocks was proposed to solve the segmentation difficulties caused by category imbalance in medical images.The potential information in medical images was fully utilized by the network combin-ing feature extraction unit with different scales.The extracted feature information was integrated and shared through dual encoder-decoder structure.The category of pixels can be better classified by the network,which can improve the accuracy of segmenta-tion results.The better segmentation results can be achieved by this method comparing with State-Of-The-Art methods on multiple medical image datasets.The experimental results showed that the segmentation difficulty caused by category imbalance can be overcome effectively by this method.Secondly,a full convolutional neural network basing on the space and channel at-tention residual block was proposed to solve the problem of noise interference in medi-cal images.On the one hand,attention can enhance the classification ability of features in the feature extraction stage,which can strengthen intra-class consistency of features and reduce noise interference in segmentation tasks.On the other hand,higher-level fea-tures can be used to direct the lower-level features to further suppress the expression of irrelevant regions during the jump connection stage.Moreover,the classification abil-ity of distinguishing the pixel category of the network was enhanced by combination with multiscale blocks and dual encoderdecoder structure.The experimental results on multiple datasets showed that the proposed method can effectively overcome the noise interference and improve the robustness.Finally,the multimodal medical image segmentation network basing on image fu-sion was proposed to improve the information utilization in multimodal medical images.The multimodal medical images were mapped to the same semantic space for fusion through fusion network.Effective integration of multimodal medical image informa-tion can be realized.Later,varieties of information was integrated into the segmentation network to improve the segmentation performance of the network,which can get more refined segmentation results.In addition,the space and channel attention residual block was introduced into the segmentation network to realize the correct classification of pixel category of network and overcome noise interference.The experimental results on multimodal glioma datasets showed that the proposed method can accurately seg-ment gliomas in multimodal medical images.