Research On Medical Image Segmentation Algorithm Based On Deep Neural Network

Medical image segmentation is an important basis for medical image analysis.A good medical image segmentation algorithm can effectively improve the performance of medical diagnostic.Existing medical image segmentation usually relies on the experience of doctors and experts,and cannot meet the needs of large-scale medical image analysis.In this thesis,we explore the segmentation algorithms of two kinds of typical medical images,medical glandular images and retinal blood vessel images.Among them,the gland is an important human histological structure,and the segmentation of glandular objects plays a key role in adenocarcinoma grading as the shape of the glands reveals the pathologic condition of adenocarcinoma;Meanwhile,the state of the human retina vessels,which are slender and long,usually reflects the healthiness of human eyes,and the segmentation of the issue is an important factor in the diagnosis of ophthalmic diseases.Due to the complexity of the two human tissue structures and the diversity presented in medical imaging,existing researches have made certain progress but still been far from perfect.To this end,we make further investigation based on previous studies,and the works are as follows:(1)A unified two-parallel-branch deep neural network for joint gland contour and segmentation learning is proposed.The glands have complex and clustered structure.In the task of gland segmentation,the contour prediction of the glands can usually help to obtain better boundaries.Existing methods usually extract different high-level features from shared low-level layers in a deep framework for separately learning gland segmentation and contour prediction and fusing the results.Such an architecture does not fully respect the complementary relationship between the two tasks,and the independency between the two kinds of task-specific features,which are meant to depict different parts of gland objects.To address the issues,we propose here a new unified end-to-end trainable deep neural network.It consists of two parallel branches,each extracts high-level features from separate low-level feature maps for a specific task under deep supervision.The gland segmentation and contour learning are jointly performed based on combined features of the two branches,while their correlations are explored through feature propagation.Besides,the proposed architecture better facilitates leveraging the power of transfer learning,which alleviates the quandary of insufficient training data and eases the learning process by weight migration from multiple task specific pre-trained models.Experiments on the benchmark dataset of 2015 MICCAI Gland Segmentation Challenge show that the proposed method delivers superior performancein the tasks of gland detection,segmentation and boundary prediction over the state-of-the-art approaches.(2)A RW-Unet deep framework is proposed by embedding a trainable random walker in the U-Net for retinal blood vessel segmentation.The two-branch network proposed previously suits for segmenting the objects that have notable appearance differences between internal structure and boundary,so it cannot be applied to segmenting the retinal blood vessel that is slender and long.To address the issue,this thesis studies the classic U-Net network,and proposes a RW-Unet retinal blood vessel image segmentation framework that integrates trainable random walk layers.The traditional U-Net performs segmentation prediction by aggregating multi-scale feature encoder-decoder units.It ignores the correlation among the foreground pixels,and the predicted labels are not locally coherent.In RW-Unet,the random walker network layer is introduced to quantify the correlation between adjacent pixels,which is further used to constrain the labelling consistency in foreground area during network optimization.It alleviates the problem of fragmentation and boosts the segmentation performance.A comprehensive experimental comparisons are made on the retinal vascular segmentation standard data set DRIVE.The results show that the proposed method can achieve good segmentation performance and demonstrate the effectiveness of the algorithm.