Research On Brain MRI Segmentation Based On Deep Feature Transfer

Magnetic Resonance Imaging(MRI)is a non-invasive imaging technique,which can be a good indication of the body tissue.Due to its safety,high speed and high spatial resolution,it is widely used in clinical analysis and diagnosis of brain diseases.With the development of computer vision,the semantic segmentation technology has become one of the hot spots in the field of human brain image analysis.Further more,deep learning-based segmentation methods developed extremely fast.By extracting deep features from a large amount of data,deep learning-based methods can automatically find useful feature representations to learn the relevant knowledge from brain MRI.Thus,it shows a good robustness in blurred boundaries,low contrast,noise and so on.However,the existing deep learning methods still have the following three problems in brain MRI segmentation:(1)In brain MRI data,there is a large difference between the number of hard and easy voxels,and the existing deep network typically performs poorly in the imbalanced data.(2)Due to the large gap between the training set(source domain)provided by medical institutions and clinical data(target domain),the model we trained on the source domain often performs poorly in the target domain.Existing unsupervised domain adaptation methods cannot bridge this large cross-domain gap.(3)The existing augmentation-based segmentation methods have insufficient generalization ability on small-sample brain MRI datasets.Based on feature transfer in deep learning,this thesis focuses on these three problems in brain MRI segmentation.The main research includes the following three aspects:1.Aiming to solve the problem of hard-voxel imbalance in brain MRI segmentation,Deep Feature Transfer-based Segmentation Network(DFTS-NET)is proposed.On the one hand,the deep features of difficult samples are transferred to enhance the network’s ability to discriminate the features of difficult samples.On the other hand,the diversity of difficult samples is increased through the intra-class transfer deformable network,thereby improving the learning ability of the segmentation network for difficult samples.In addition,Regionrelated Focal Loss(RFL)is proposed to solve the problem of hard-voxel imbalance by weigh those hard-classified voxels.Specifically,based on the defined focalized point,we reshape the standard Dice Loss(DL)by down-weighting the loss of well-classified voxels and up-weighting that of hard-classified ones with a dynamic logarithmic factor to solve the hard-voxel imbalance.The experiments show that in the segmentation task of normal and abnormal brain tissue,the proposed DFTS-NET can classify hard-voxels more accurately,which effectively improves the segmentation accuracy for hard samples.2.Aiming to solve the problem of existing methods’ misalignment of inter-domain features due to inter-domain differences in cross-domain segmentation of neonatal brain MRI,an joint Intra-classly Adaptive GAN and Segmentation Network(IAS-NET)is proposed.Specifically,intra-class adaptive GAN(IA-NET)is a 2D generative method.Through the proposed Local Adaptive Instance Normalization(LAda IN),it can effectively transfer the intra-class deep features of the target domain to reduce the feature misalignment between domains,thereby enhancing the robustness of the model in the segmentation of target domain images.In addition,in order to improve the segmentation accuracy of important regions in neonatal brain development and simplify the clinical application,a Joint Deformable network and Intra-class Transfer GAN(DIT-NET)is proposed.Specifically,intra-class transfer GAN introduces a 3D LAda IN algorithm in Unsupervised Domain Adaptation to transfer the intra-class features of target domain images.The deformable network enhances the shape consistency between the synthesized image and the target-domain image by learning the cross-domain shape difference.The 3D segmentation network aims to achieve end-toend cross-domain segmentation.The experiments show that,compared with the existing methods,the proposed method can more effectively overcome the intensity and shape differences between the different domains,which improves the segmentation performance of the segmentation model in the target domain.3.Aiming to solve the problem of small-sample dataset problem in brain MRI segmentation,a Random Transfer GAN-based Small Sample Segmentation Network(RTSNET)is proposed.Specifically,through the improved LAda IN algorithm,the random transfer GAN(RT-GAN)can learn the intra-class distribution differences between samples in a limited dataset.By extracting deep features with random deviations during transfer,RT-GAN can increase the diversity of features and create samples with large diversity.The multiple deformable models can generate samples with different degrees of deformation by learning the shape difference between samples,thereby increasing the shape diversity in the synthesized dataset.The experiments show that the proposed RTS-NET can effectively improve the generalization of the segmentation model in the small-sample brain MRI dataset,which improves the accuracy of the segmentation results.