Research On Intelligent Recognition Method Of Liver Cancer Based On Multi-modality Non-contrast Mri Image

Liver cancer is one of the most common and high mortality cancers in the world.It is difficult to detect the tiny tumors of liver cancer in time,which is the main reason for the high mortality.At present,magnetic resonance imaging(MRI)is one of the main clinical diagnostic methods for liver cancer.It can significantly improve the sensitivity and specificity of tumor diagnosis by injecting contrast agents.However,the injection of contrast agents inevitably has problems such as high risk,high toxicity and large side effects.In recent years,non-contrast agent MRI(NCMRI)has become a new focus of research due to its short scan time,low risk,and no toxic metal hazards.However,the use of non-contrast scanning will make the contrast of the MRI image lower and the edges of the tumor area blurred,which adds many difficulties to the accurate diagnosis of liver tumors.Aiming at this bottleneck problem,it is planned to use a computer to perform intelligent diagnosis of liver cancer on NCMRI.This will not only become an auxiliary tool for clinicians,but also an important method to improve the survival rate and cure rate of patients.In this thesis,a series of researches on the intelligent diagnosis of early liver cancer based on NCMRI are carried out.Through the analysis of NCMRI images,the 2D segmentation method of tumors is researched with the help of artificial intelligence technology;on the basis of fully studying the complementary information between multi-modality NCMRI,explore the construction of 3D segmentation and multi-index quantification of liver tumors Multi-task correlation network;mining the dependent information among multi-modality NCMRI images,and at the same time constructing a small sample data enhancement system to realize the automatic classification of tumors in liver NCMRI;further study the relationship between segmentation and detection tasks Correlation,construct a joint adversarial learning network for liver tumor segmentation and detection.The main research contributions are as follows:(1)Aiming at the characteristics of blurry edges of tumors in NCMRI,low contrast with surrounding normal tissues,large differences and high complexity,this thesis proposes a Radiomics-guided DUN-GAN network for 2D liver tumor regions segmentation.The research content of this part includes three aspects: i)The segmenter of density nested UNet extracts the semantic features and global guidance features of liver tumors,and uses the feature correlation to fuse the features reasonably and efficiently.ii)Radiomics features combined with three-phase contrast-enhanced MRI(CEMRI)(i.e.arterial phase,portal venous phase,and delayed phase)are added into the discriminator as prior knowledge to learn the mapping relationship between the segmenter and the discriminator.iii)Under the guidance of the idea of radiomics feature confrontation,the network obtained a highly accurate 2D segmentation of the liver tumor in the NCMRI.The prior knowledge of radiomics guidance and the multi-strategy fusion method proposed in this thesis can extract the key hidden radiomic features of the contrast agent in the non-contrast agent image,so as to realize the more accurate liver tumor segmentation in the low-contrast image.(2)In view of the complex spatial anatomy of the tumor in 3D NCMRI,the large differences in tumor representation among multi-modality NCMRI(i.e.T1FS+T2FS),and the high complexity of completing multiple tasks(i.e.segmentation + quantification)at the same time,etc.This thesis proposes a two-stream associated 3D(Two-stream associated 3DNet,TsA-3DNet)heterogeneous learning network,which uses multi-modality NCMRI to simultaneously perform 3D segmentation and multi-index quantification of liver tumors.The research content of this thesis mainly includes two aspects: i)The network fully considers the relevance and difference of multi-modality NCMRI,and the proposed multi-view fusion mechanism(i.e.Multi-view fused,Mv F)uses a proportional gate optimization algorithm to extract multimodality data.Complementary information to further integrate multi-channel(i.e.T1FS+T2FS)and multi-scale features.ii)The network fully combines the associations and constraints between multiple tasks,and the proposed bi-directional guided(Bd G)mechanism uses the bidirectional guided mechanism to correlate the complementary information between multiple tasks to further realize the simultaneous multi-task(i.e.segmentation + quantification)optimization.The heterogeneous learning network proposed in this thesis can establish a multiview feature fusion and multi-task bidirectional guidance mechanism,so as to achieve more accurate 3D segmentation of liver tumors in multi-modality images and comprehensive multiindex quantification at the same time.(3)Aiming at the problems of few NCMRI samples,higher auto-recognition complexity,and lower tumor specificity,this thesis proposes a stable detail-guided generative Adversarial network(Dg-GAN)network to improve liver tumor classification performance by enhancing data.The multi-scale perception generator in Dg-GAN introduces the residual block perception context information,and strengthens the extraction of detailed features to synthesize more realistic contrast-enhanced images.In this thesis,convolution and global average pooling are performed on the synthetic data-enhanced data set,and a detailed-guided VGG model is designed to verify the effect of data enhancement on the classification performance of liver cancer.The detailed feature map of the generator is used as a guide to introduce the classifier to facilitate the extraction of specific features of different types of liver cancer and improve the performance of tumor classification.The data enhancement method proposed in this thesis effectively overcomes the problem of fewer data samples and demonstrates better automatic classification performance.(4)Aiming at the problem of large feature differences between multi-modality NCMRI and high complexity between segmentation and detection tasks,this thesis proposes a united adversarial learning framework(UAL)network,which integrates complementary information of multi-modality NCMRI realizes the simultaneous segmentation and detection of liver tumors in a joint constraint manner.The encoder enhances the key feature extraction of the three modalities by introducing the prior knowledge of the edge difference feature map,and at the same time uses the gate mechanism to achieve feature fusion and adaptive selection.The coordinate sharing mechanism enables segmentation and detection to be combined to achieve unified learning,and at the same time,multiphase radiomics features and semantic features are combined to achieve segmentation and detection while adversarial learning.The joint adversarial learning network proposed in this thesis can sufficiently improve the performance of liver tumor segmentation and detection.In summary,this thesis focuses on the four problems of the intelligent diagnosis of liver NCMRI,and carries out the research of related tasks and the construction of the network from different purposes.This work has certain computer theoretical significance and auxiliary diagnosis value.