| To solve the low medical efficiency existing in the medical and health system,the development of intelligent medicine has become an inevitable choice.As the basis of various medical image processing tasks,medical image segmentation is of great significance for the realization of intelligent medicine.However,due to the problems of small datasets,imbalanced sample distribution,multiple models,low contrast and large noise interference,the accuracy of medical image segmentation using deep learning algorithm cannot meet the clinical requirements.Therefore,this thesis focuses on two mainstream segmentation tasks,namely,retinal vessel segmentation and liver segmentation.This thesis modifies the structure of UNet++,not only improves the segmentation accuracy of the network,but also lays a foundation for the clinical use of medical image segmentation.The contributions of this thesis can be summarized as:1.An optimization method for the feature extraction ability of the encoder is proposed.The Transformer structure is introduced into the encoder to supplement the global context information and the ablation experiments and comparative experiments on the DRIVE dataset show that the proposed method can effectively improve the ability of retinal vessel segmentation.2.A new feature pyramid attention model is proposed to optimize the feature fusion mode of the decoder,and the ablation experiments and comparative experiments on the DRIVE dataset show that the proposed method can improve the performance of retinal vessel segmentation.3.An adaptive weight loss function is proposed.Considering that the deep supervision of the current network ignores the different feature expression ability of networks with different depths,the horizontal and vertical adaptive weight loss functions are designed to assist the network fitting.The results of comparative experiments on the LiTS 2017 dataset suggest the effectiveness and the universality of the proposed method in improving the segmentation effect of liver tumors. |
