Interactive Image Segmentation Algorithm For Medical Images Research And Application

Medical image segmentation is a technology that has been flourishing since the beginning of the new century.The task of medical image segmentation has become an area of interest for many researchers as it requires preliminary processing of medical images in many medical tasks.In order to obtain better segmentation results many classical algorithms and networks have been proposed,such as machine learning algorithms,deep learning and even graph theory algorithms.In thesis,we focus on the application of interactive methods in deep learning in the field of medical images.By combining interactive methods and deep learning in medical images,interactive methods are used to guide the optimization of network parameters toward user-specified effects by providing interaction information to the deep learning framework.In this way,medical images can avoid the problems of incomplete segmentation and imprecision in interactive segmentation caused by deep learning.In the thesis,we address the problems of interactivity in medical images,so as to improve the structure of the interactive framework to make it more suitable for real task scenarios and improve efficiency.The research in thesis includes:(1)interactive image segmentation under joint training of multiple networks.In traditional deep learning,the fitting is done by the parameters and structure of the network,which is limited by the network structure,so there are often many network structures to handle the same task,which will cause the effect of over-redundancy.By using multiple networks trained jointly and interacting with each other,it can help the user to quickly locate the location of the segmentation error and thus correct this error.(2)Multi-scale information interaction image segmentation.In most of the interactive segmentation networks,users interact with segmentation results that do not contain the spatial information of the images.Such segmentation results do not reflect the use of spatial information of the image by the network,and therefore the user cannot interact with such information.By using the multi-scale information interaction network segmentation method,the user can interact with the multi-level extraction features of the network thus affecting the overall extraction feature capability of the network and enhancing the final segmentation results.(3)Interactive optimization for the tiny structure of medical images.In many medical images,there are many tiny objects that are difficult to segment,and even after using interactive to improve network segmentation accuracy,it is difficult to achieve perfect segmentation of these tiny objects,so this thesis proposes a region growth algorithm to optimize the interactive method to solve these problems and achieve detection of tiny objects.