| Computer-aided diagnosis technology is widely used in clinical practice,and medical image segmentation is one of the key steps.Its purpose is to segment the targets with certain special meanings in medical images and extract relevant features to provide a reliable basis for subsequent clinical diagnosis and pathological research,and assist doctors to make more accurate diagnosis.However,the localization and segmentation of targets in medical images is still a pressing technical problem due to the different morphologies and complex structures of different medical image segmentation targets,and the difficulty of distinguishing between some imaging techniques that image different organs and tissues with similar intensity,color or texture features.The main task of this paper is to study the medical image segmentation algorithm based on deep neural networks,which can accurately segment the target tissues from medical images and provide assistance to doctors for diagnosis and further research.This paper presents an in-depth study of recent medical image segmentation methods based on deep neural networks,introduces their basic principles and their respective contributions,and discusses in depth the characteristics of medical images and the difficulties of segmentation.Combining some characteristics of medical images,this paper proposes three different medical image segmentation algorithms for different segmentation task characteristics to improve the segmentation effect of specific medical images.The specific research results and contents are as follows.(1)A Transformer-based 2D medical image segmentation network is proposed for the 2D medical image segmentation task.In order to solve the problem that convolutional neural networks cannot learn the connection between remote features well,a Transformer-based Ushaped network architecture is proposed,and a branching structure based on convolutional neural networks is additionally designed to alleviate the problem that the pure Transformer network structure is difficult to extract local detail features of images.A point sampling mechanism is added in the upsampling stage of the network to refine the segmentation edge prediction to alleviate the problem of blurred medical image edge segmentation caused by oversampling.Finally,a hybrid loss function is designed to reduce the impact of the positive and negative sample imbalance problem of medical images on the segmentation effect.(2)For the CT liver image segmentation task,an improved full-scale jump-connected network structure is proposed to improve the overall network’s focus on the segmentation target region by designing an attention gate module between the shallow encoder and the deep decoder to learn a scale factor for the fusion of the two features.In addition,based on the traditional point sampling mechanism,an improved point sampling strategy is proposed for the characteristics of medical images,which is used to further improve the edge segmentation effect of CT liver targets.(3)For the task of segmenting small targets in medical images,a different idea from the traditional medical image segmentation is proposed.The RPN network is first used to roughly frame the region where the segmentation target is located,and then the FSRCNN network is used to super-resolve the frame region to achieve the purpose of transforming the small target segmentation task into a large target segmentation task,and then a shallow U-shaped medical image segmentation network is used for segmentation.Finally,all segmented targets are mapped to the same region of the original image to obtain the final segmentation results.The above methods are all experimentally analyzed and comparatively validated on four publicly available medical image datasets,CHAOS,GLAS,DSB18 and Mo Nu Seg.By comparing different evaluation metrics,the results prove the effectiveness of the proposed method in this paper. |
