Optimization And Implementation Of Medical Image Segmentation Algorithm Based On Deep Learning

Malignant tumors have been one of the main diseases that endanger human health,and their morbidity and mortality have remained high.In the course of radiotherapy and chemotherapy for treating tumors,the positioning of the target area of the tumor area and the segmentation of important surrounding organs have always been important.Previously,the manual segmentation performed by doctors was huge.How to further reduce the workload of doctors and improve the efficiency of treatment has been a hot research topic in recent years.With the rapid development of artificial intelligence technology,deep learning algorithms have made breakthrough applications in computer-aided diagnosis.The image segmentation algorithm based on deep learning is applied in the medical field,which can quickly realize automatic target area delineation and organ segmentation,alleviate the shortage of medical resources,and improve the rate of medical treatment for patients.Significant for the treatment of malignant tumors.This thesis mainly researches and implements the medical image segmentation algorithm 3D-unet based on deep learning.The main contributions are as follows:1.Proposed a horizontal depth multi-scale U-shaped convolutional neural network(Level Depth Multiscale U-net,LDM-Unet)segmentation algorithm,based on 3D-Unet,using a multi-convolution stacking method to extract multi-scale information in the horizontal layer To improve the network's ability to perceive targets of different scales,use hole convolution at each depth layer to extract spatial structure information at different depths,to avoid the information loss caused by continuous pooling,and thus improve the network's segmentation performance.In the public data set BraTS 2018,the division dice coefficients of WholeTumor,TumorCore and EnhancingTumor were respectively 0.82,0.73 and 0.64.2.A multi-resolution parallel U-net(MR-Unet)segmentation algorithm is proposed to maintain multiple subnets with different resolutions,and the subnets are connected in parallel.Method and intensive information exchange,avoiding the problem of inaccurate spatial position caused by traditional serial networks to recover resolution through upsampling.Segmentation dice coefficients of WholeTumor,TumorCore,and EnhancingTumor on the public dataset BraTS 2018 have obtained results of 0.84,0.77,and 0.71,respectively.3.Applying the above model algorithm to the West China Celiac dataset,a manual feature extraction fusion enhancement method is proposed to perform different feature extraction on CT image data to increase the number of channels in the image and increase the information capacity in the image.Expansion of the data.A new weighted loss function is used to solve the problem of class imbalance in image segmentation,and the algorithm is optimized and implemented in engineering to achieve the outline of clinical target areas and organ segmentation of rectal cancer.