Research And Implementation Of Multi-organ Segmentation Algorithm Based On Deep Learning

Organ segmentation is a key task in medical images,and it is also the basis for tasks such as computer-aided diagnosis(CAD),diagnostic intervention,and disease treatment and rehabilitation planning.In some radiotherapy plans,it is necessary to accurately outline the target organs in the medical image in advance.The organ labeling process requires a lot of labor and is very cumbersome,so it is easy to introduce human errors.Multi-organ segmentation refers to the use of the same model to segment multiple organs at the same time,which is more challenging than single organ segmentation.On the one hand,the large size difference between different organs makes it easy for data-driven deep learning models to ignore the characteristics of small organs with fewer voxels.On the other hand,the morphology and structure of different organs are quite different,and even the differences between different patients of the same organ are also very obvious,which puts forward extremely high requirements for the robustness of the deep learning model.The current deep learning models in medical image tasks are mainly neural networks represented by FCN,U-Net,V-Net and their various variants.Cascading two identical networks is a very popular method in both research and engineering fields.It uses the first network for rough feature extraction,and then uses the second network for further refinement,which can significantly improve Network performance.This article is based on the cascaded V-Net network,using the Seg THOR 2019 competition and the Multi-Atlas Labeling Beyond the Cranial Vault(MALBCV)competition data set to conduct experiments,and do the following four points of work and research:(1)Establishing jump connections between corresponding modules at the same level in the cascade network greatly improves the performance of the cascade structure.The traditional cascaded network usually only uses the final output of the first network,analyzes it and extracts the required information,and sends it to the second network,but ignores the learning in each module of the first cascaded network To the characteristics.In this paper,by establishing a block-level jump connection between the two network corresponding modules,it makes full use of the useful features learned by the first cascade network and enhances the performance of the cascade structure.(2)In the neural network,the large convolution kernel and the small convolution kernel are mixed to enhance the feature learning ability of the network.Generally,the model using a small convolution kernel is smaller in size and has better performance than a large convolution kernel,but a large convolution kernel has more advantages in capturing information between distant voxels.After the mixed use of the two in the network,this paper has made a significant improvement in dealing with tasks with large differences in size such as multi-organ segmentation.(3)Some small organs in multi-organ segmentation,such as esophagus,have long and narrow shapes and obvious morphological differences between different patients.The segmentation effect is usually poor.Therefore,by cutting and using a single classification network for separate segmentation,the segmentation effect of such organs is greatly improved,making the segmentation of such difficult-to-segment organs more practical.(4)Taking into account the security issues in the medical field,this article uses FGSM,BIM,PGD,and CW four classic anti-attack methods to test the anti-attack ability of the network.The experiment found that the current commonly used networks are difficult to resist white box attacks from the four attack methods,which is also a direction worth exploring in the future.