Research On 2.5D Medical Image Segmentation Based On Improved U-Net

Medical images are crucial for the accurate diagnosis of diseases,and image segmentation is a key step in analyzing medical images.In recent years,computerassisted medical image segmentation has typically been performed using 2D segmentation networks,which cannot capture the spatial information contained in 3D images.As a result,3D segmentation networks have become a research hotspot.However,3D segmentation networks require higher hardware configurations and more computing resources.In addition,a large number of parameters and computational costs limit the design of deeper and more complex network structures.In this thesis,we propose a 2.5D segmentation method based on an improved U-Net network,with the goal of seeking a balance between segmentation accuracy and computational cost to meet clinical application requirements.The main work of this thesis is as follows:1)In terms of system construction,this study conducts research on relevant theories and compares the performance of two popular networks,2D U-Net and 3D U-Net,through experiments.The experimental results show that 2D U-Net has the advantage of high processing efficiency but cannot capture spatial information,while 3D U-Net has the advantage of processing three-dimensional images but has a high computational cost.In addition,the insufficient processing of local details caused by the use of the traditional U-Net network structure further limits the improvement of segmentation accuracy.2)A full-scale fusion-based 2.5D segmentation method is proposed to address the deficiencies of traditional 2D and 3D segmentation networks.The proposed method is based on a solution that incorporates the strengths of both approaches.Firstly,the raw 3D image is utilized to obtain slice data along three orthogonal axes,and these are then inputted into an improved U-Net model that incorporates full-scale connections.The resulting predicted outputs from each axis are merged with equal weights to produce the final output.This method preserves the efficiency of 2D networks while capturing spatial information,resulting in accurate 3D segmentation results with reduced computational costs.To evaluate the proposed 2.5D segmentation method’s performance,two representative medical datasets,Li TS17 and Bra TS20,were selected for CT and MRI,respectively.Experimental results demonstrate that the proposed 2.5D segmentation method outperforms traditional 3D segmentation networks in terms of Dice,IoU metrics,and network computational cost(Param).3)A 2.5D segmentation method based on an improved encoder is proposed.This method addresses the insufficient local detail processing caused by the use of the traditional U-Net network structure and proposes a solution.Firstly,the FRN normalization layer is used instead of the traditional BN layer to reduce the influence of batch size on the network.In addition,the residual structure is fused in the encoding part to improve the feature extraction ability of the encoder.Furthermore,deformable convolution is integrated into the convolution process to adaptively adjust its receptive field according to the input features,thereby capturing more complex and subtle image details.Experiments were conducted on the publicly available liver CT dataset(Li TS17)and brain tumor MRI dataset(Bra TS20).The experimental results showed that compared with the traditional 2D and 3D segmentation methods and two other advanced 3D segmentation methods,The 2.5D segmentation method in this paper has better performance in Dice,IoU index and network computing cost.