Research On Medical Image Segmentation Technology For Tiny Tissues

The purpose of medical image segmentation is to accurately outline the contour of human organs,lesions or tissues.As a prerequisite for common disease screening and lesion analysis,it is an important step for accurate diagnosis.However,it takes a huge cost to segment the medical image with the manual method.Therefore,under the premise of the imbalance and insufficiency of medical resources in China,it is of great significance to use computer technology to segment the medical image automatically.In this way,it can greatly relieve the working pressure of medical personnel and provide auxiliary guidance for clinical diagnosis.Due to the characteristics of medical images,the segmentation of micro structures is a common task requirement,such as fundus vessel segmentation,tissue cell and optic disc and cup segmentation.Starting from these demand scenarios and combined with the research status recently,this thesis has carried out the following three research points:Segmentation on fundus vessel and cell via self distillation method.In these two tasks,the target structures are numerous and densely distributed,which makes it difficult for the existing methods to accurately segment the vessel endings and cell boundaries.Aiming at this phenomenon,this thesis proposes a self distillation training mechanism that can be embedded into most encoder decoder networks,and experiments are carried out on variable public datasets.The experimental results show that the training method can improve the performance of network segmentation and has strong universality.Optic disc and cup segmentation network based on enhanced sampling module.Aiming at the phenomenon of information loss during upsampling in the current methods,this thesis analyzes the causes of this phenomenon,and proposes a segmentation network based on enhanced sampling module.Experiments on multiple datasets show that the proposed network can generate accurate disc and cup prediction images.In addition,the advancedness of the method proposed in this thesis is verified by comparing with other methods.Segmentation optimization method based on self-supervised learning.Finally,this thesis introduces a self-supervised method for optimization in response to the sparse data volume in medical image datasets,which leads to the phenomenon that the network is easily overfitted during training and thus the image segmentation effect.In the experimental section,a self-supervised method that best suits the needs of medical image segmentation tasks is selected by comparing and analyzing different upstream tasks and parameter migration methods,and it is demonstrated that the self-supervised approach can improve the network segmentation performance by using easily collected unlabeled data.