Research On Semantic Segmentation Of Multi-Distributed Objects And Small Objects For 3D Medical Imaging

Medical imaging is the primary source of information for disease screening,diagnosis and treatment.Using computer-aided diagnosis technology to further intelligently analyze and mine medical image information to assist doctors in interpreting medical images has become an important requirement for the development of modern medical imaging technology.In recent years,with the rapid development of artificial intelligence,machine vision and other technologies,deep neural network models have performed well in image visual recognition,providing new ideas for medical image analysis,and have been widely used in medical image segmentation tasks.The use of deep neural network models to automatically segment medical images can improve the efficiency of diagnosis and treatment.So it has attracted the attention and research of many researchers.However,automatic segmentation of medical images remains a challenging problem even with deep learning.Existing methods have also failed to achieve satisfactory results.The reasons can be described from four aspects:（1）The contrast between the target area and surrounding organs and tissues in medical images is low,and the anatomical structure is complex.The accuracy of image segmentation methods is low;（2）the target regions in medical images have the characteristics of size and shape diversity,and a large amount of semantic information will be lost simply by downsampling operations,and it is difficult for the network to model multi-scale features in this task;（3）Compared with natural images,the resolution of 3D medical images is larger,the existing attention mechanism is easy to introduce more training parameters,and needs to bear a greater computational burden;（4）The lesion region in medical images with the characteristics of multi-distribution,the existing regional-level-based loss functions are difficult to accurately segment multi-distributed lesions.Aiming at the problem（1）,we design a Multi-Structure Response Filter（MSRF）to extract the prior information of the 3D structure.In contrast to traditional methods,MSRF is designed to enhance other tissues and organs with specific geometries rather than the structures to be segmented.Therefore,the enhanced image can guide the segmentation network to distinguish between the structure to be segmented and the similar organ tissue in the image.For problem（2）,we design a multi-scale contextual feature extraction module to fully capture contextual information.Specifically,we use atrous convolution in the convolution block of U-Net to extract multi-scale features of target regions.In addition,due to the small number of medical image datasets,models with a large number of trainable parameters are prone to overfitting,and excessive use of various network components for feature extraction and fusion is not suitable for this task.Therefore,we introduce residual connections on the basis of atrous convolution to achieve more effective multi-scale feature learning.For problem（3）,we utilize the 3D structural prior information to guide the execution of the spatial attention mechanism in the Prior Attention Module（PAM）,which facilitates the recovery and fusion of key spatial information.Our PAM can effectively model the importance of different regions on medical images instead of learning from feature maps,which helps the model to extract features related to target regions.In addition,in order to reduce the amount of parameters in the attention mechanism and improve the ability of model to model the channel domain and the spatial domain,we propose the Shift-Channel Attention Module（S-CAM）.The S-CAM uses the strategy of shifting some channel features to model the feature relationship in adjacent channels,and enhances the interaction of features between adjacent channels without introducing additional training parameters,thereby improving the discriminative features.expressive ability.For problem（4）,we propose a novel loss function called Weighted-Region（WR）loss function.The WR is an improvement over existing region-level based loss functions.Existing region-level based loss functions directly compute the sum of multiple regions,while our WR loss function can implicitly learn the weights among multiple regions without manually specifying the weights.In the segmentation task of multi-distributed targets,the goal of WR is to increase the weight of small target regions and alleviate the problem that the segmentation of small target regions is difficult to be further optimized when large target regions are in the majority.Finally,we validate all the proposed methods on three datasets,including a private esophageal cancer dataset,two public Li TS liver tumor datasets and the 3Dircadb liver tumor dataset.In the esophageal cancer segmentation task,the Dice Similarity Coeffificient（DSC）for esophageal cancer segmentation was 84.839%,the Precision was85.955%,the Sensitivity was 83.752%,and the Hausdorff Distance（HD）was 2.583 mm.In the liver and tumor segmentation tasks on the Li TS dataset,the DSC for liver segmentation was 96.9%,the Volumetric Overlap Error（VOE）was 2.56%,and the Average Symmetric Surface Distance（ASSD）was 0.96 mm.The DSC for tumor segmentation was 75.1%,the VOE was 21.09%,and the ASSD was 1.08 mm.Furthermore,we utilize the 3Dircadb dataset to test the generalization performance of the model,with 96.47% DSC,9.68% VOE,and 2.98 mm ASSD for liver segmentation.The DSC of tumor segmentation was 74.54%,the VOE was 34.42%,and the ASSD was1.21 mm.Experimental results show that the proposed method outperforms the existing state-of-the-art networks in all three datasets.