Study On Medical Image Segmentation Key Approaches Based On Multi Models Fusion

With developments of medical imaging devices,related critical techniques and computer skills,medical image analysis plays an important role in clinical diagnosis and therapy.Medical image segmentation has a close relationship with other links and performance of the entire system.However,most traditional algorithms serve an acceptable performance in natural images rather than medical ones,as a result of strong noises widespread in the medical image.Besides,the foreground and background provide blurring boundary,low contrast,abnormal anatomical structure,irregular shape and heterogeneous intensity texture and etc.In recent 20 years,although level set models made remarkable progress,they belong to a semi-automatic algorithm and are trapped into these challenges.Recently,convolutional neural networks based on deep learning succeed to be applied in medical image segmentation with an improved performance,but they face many problem s still,for example the high computing complexity and reliance on the-scale of training set with pixel-wise annotated level.At.present,researchers try to use multi·models fusion to resolve inherent issues of algorithms in medical image segmentation This strategy is able to help different m ethods with complementary advantages generating advanced results.This thesis focuses on the study on key approaches of medical image segmentation based on multi models fusion.Main research contents and innovations are as follows:(1)To address the issues of over-segmentation and under-segmentation via traditional level set model on different liver tumor CT images,this thesis proposes a novel dynamic parameters estimation level set model using 3D corivolutional neural network.At first,this thesis presents a 3D convolutional neural network that predicts probabilities of evolving contour position relative to lesion boundary.To resolve traditional level set model sensitive to different initializations,this thesis uses 3D convolutional neural network to choose the candidate close to lesion boundary,which prompts the automatic level further.In addition,this thesis has the research of dynamic regulation on control parameters of level set energy functional via 3D convolutional neural network outputs,strengthening the adaptiveness of model on different heterogeneous texture.Localizing regionlevel set holds the local window for complex shape better,and if the size is constant this mechanism will restrict the evolution of level set model.This thesis gives a new scheme to regulate the window size dynamically using the probability output,which cannot result in computing resources burden for 3D network.Sufficient experiments indicate that,compared with traditional level set model,the proposed method serves obyious predominance on liver lesion 3D segmentation.On the dataset with high specificity,our proposed model has the best fitting ability,especially for the lesions located close to boundary with great robustness.Then compared with the state-of-the-art algorithms based on machine learning or deep learning,the proposed method outperforms others in all metrics,which demonstrates the proposed level set model combing deep learning skill can be adaptive automatically different lesions with distinct features,which ensure.the energy functional of being optimized to global minimum.Furthermore,it has capability of resistance to strong noises,extremely reflecting advantages ’of level set model on irregular shapes of lesions and providing notable generalizations in other areas.(2)For the problem of 3D convolutional neural network with excessive computing costs and false positives for 3D segmentation,this thesis proposes a multi-task 3D convolutional neural network with an end-to-end trainin g process At first,this thesis designs a multi-task network with encoder-decoder symmetrically.The encoder branch is responsible for salient detection.For strengthening detection,an attention activation block is introduced to localize the object accurately,the mechanism of which is light and cannot produce huge costs for 3D model After acquiring localization information,this thesis provides a strategy of estimation to the scale of 3D interesting region reducing computation cost of 3D neural network..In addition,this thesis devises a cropped feature fusion bridging encoder and decoder branches.In th e procedure of down-sampling and up-sampling.via encoder and decoder this design can extract.and fuse feature m aps in different hierarchies with resolutions,which helps not only salient detection but serves details in the end of the network.At last,this thesis designs a hybrid loss function focusin g on blurring boundary in training process role.With multi-resolution models integration,this m odel achieves automatic 3D segmentation of maxillary sinus.For demonstration of the proposed model’s superiority,this thesis did sufficient experiments to compare with the state-of-the-art method for 3D segmentation.The results show that the proposed method has an obviously more advanced performance,which gives a proof that the region of interest estimation strategy cannot only reduce computation cost of 3D model but eliminate false positives because of a more significant shrink region estimated,promoting models based on 3D convolutional neural network applied in practice.(3)To resolve the problem that models based on deep learning for medical image semantic segmentation rely on the scale of pixel-wise annotations greatly,this thesis proposes a convolutional neural network for semantic segmentation with semi-and weakly supervised learning to get rid of the restriction.At first,we introduce the state-of-the-art convolutional neural network for segmentation trained to a basic ability of discrimination by a limited pixel-wise annotated database.In addition,this thesis designs two branches for model fusions and implementation of semi-and weakly supervised learning.One of branches is deep level set that gives a proper initial contour by class activation mapping produced by segmentation network at the’ beginning.Different from traditional level set,the probability map produced by segmentation network joins in iterative process of energy functional to update control parameters dynamically,which improve the perform ance of level set model in noisy,low contrast,blurring boundary and heterogeneous texture.Similarly,another branch focuses on conditional random field model for segmentation under unsupervised learning.T wo branches together provide pixel-wise annotations from image-level semantic training set for optimization.of basic segmentation network,which are trained respectively with two loss fun ctions in a parallel end-to-end way.Experimen ts on skin lesion segmentation show that,compared with the state-of-the-art methods based on deep learn ing,the proposed model acquires the most advanced results obviously.Especially with models based on semi-and weakly supervised learning,the proposed model also achieves the best performance indicating great superiority.(4)Exploring the optimization of convolutional neural network for semantic segmentation,this thesis proposes a post-optimizer based on generative adversarial networks.A feature piles network as the critic network is devised and segmentation model selected as the generative network.Both of them are trained in the mix-max game.As a result of the critic network that discriminates the difference between generative data and ground truth in multi-level feature hierarchies,growth in performance of segmentation model is obvious.For better training semantic segmentation network,this thesis presents a novelloss function named narrow band suppression.It gives more attentions on the narrow band region along lesion boundary where false positives appear frequently and narrow band suppression supplies more restraints on harder pixels.This mechanism shows a direct improvement on blurring boundary.Sufficient ablation studies on ISIC-2017 dataset for skin lesion segmentation demonstrate that the hybrid loss function and post-processing optimization network respectively increase the-accuracy of segmentation network obviously,serving great generalization and prqspect.In conclusion,main work in this thesis uses the strategy of multi-task fusion to address issues in medical image segmentation remarkably,which prompts segmentation skills applied in.clinical practice effectively.