Research Of Attention Based Neural Network Technology And Its Application

With the development of intelligent medical imaging devices and the improvement of computer software and hardware technology,the research and application of medical image processing technology in the healthcare field has been increasingly deep and extensive,and it has played an important role in assisting clinical diagnosis and quantitative analysis.The dissertation focuses on the attention mechanism based neural network technology and other related issue in the field of medical image processing,and utilizes the medical image processing and deep learning to conduct researches on multimodal medical images such as Ultrasound,MR and CT images.The contributions are mainly as follows:(1)Aiming at the problem of thoracic and abdominal respiratory-induced motion tracking in the ultrasound-guided radiotherapy,a tracking method that combines the feature indexing and slow feature analysis is proposed.The statistical model is employed to reduce the noise between the tracking target and the background of the ultrasound image,and the feature retrieval atlas is constructed using the feature indexing and slow feature analysis.Afterwards,the mapping in the atlas is redefined according to the characteristics of the slow feature signal.With the feature indexing,the template image can be updated while the search range is reduced to improve the tracking accuracy.In addition,this work involves repeatability experiments of image acquisition to help clinicians ensure the correctness of the scanning position during intra-fractional guidance procedure with the ultrasound guidance device and robotic arms.The experiments verify that the proposed method can obtain lower tracking error with faster speed,and it can be potentially used for respiratory-induced motion tracking in the radiation therapy treatment.(2)Aiming at challenges of the ultrasound motion tracking under complicated clinical environments,such as the change of target shape,blurring and occlusion,a novel motion tracking method based on attention-aware and fully convolutional long short-term memory networks is proposed.The observation network is introduced to focus on the local region that contains the target,the transfer learning is leveraged to extract the deep spatial features of the ultrasound image,and the attention mechanism is employed to help the model effectively search for the target-related features to predict the target position and enhance the target image.The long shortterm memory network summarizes the characteristics of deep spatial features and the regular pattern of the target motion to predict the target position of the next frame.The multi-task learning strategy is combined with the adaptive loss weighting to improve the overall performance of the model.The experiments verify that the proposed method has higher tracking accuracy with higher speed,which can assist clinicians to ensure the tracking accuracy and robustness during the fractional ultrasound-guided radiotherapy.(3)For the issue of modality incompleteness of multimodal MR images and glioma grading in clinical glioma diagnosis,the attention-aware generate adversarial network is proposed for multimodal MR image synthesis and glioma grading.The proposed model introduces edge-aware strategy to learn the detailed features to improve the quality of image synthesis,and it employs the tumor attention to help the model heuristically choose the features that are effective for glioma grading task and enhance the weights of the feature.The model constructs a common feature space,and the common feature space explicitly learns the inherent inter-modality relationships,the domain-invariant features and also the lesion-specific representations to leverage more comprehensive multimodality information.Multiple network modules are combined with the multi-task learning strategy to improve the overall performance of the model.It is verified on the glioma dataset that the method can effectively solve the one-to-many image synthesis task and improve the glioma grading performance using a single modality.(4)Aiming at the limitations of fractional radiotherapy dose quantified by CBCT to prevent radiation-induced liver disease,a gradient-based free form deformable registration method is used to correct the electronic density of CBCT images while obtaining segmentation results of the organs automatically,and the corrected CBCT is used for the dose calculation.Combining with the clinically effective dose parameters,the statistical analysis of the adjusted dose and planned dose was performed to obtain the correlation between the radiation dose and the radiationinduced liver disease.And the statistical analysis is also performed on multiple factors(inter-fractional respiratory movements,anatomical changes,etc.)that cause dosimetric variations.Through evaluating the relationship between the cumulative dose of the primary liver cancer patients with radiation-induced liver disease after radiotherapy treatment and the hepatic radiation tolerance,the results validate the proposed method can prevent radiation-induced liver disease based on CBCT images during fractional radiotherapy.