Abdominal Image Synthesis By Multi-modal Magnetic Resonance Imaging Based On Deep Learning

Magnetic resonance imaging(MRI)is very important in the diagnosis of abdominal diseases,because of its high resolution in soft tissue,no ionizing radiation,and multiparameter and arbitrary layer imaging.T1-weighted images(T1WI)and T2-weighted images(T2WI)are the two most commonly weighted images(modes)used in the conventional MRI scans,and the scan time of T2 WI is generally 4 to 10 times longer than T1 WI.Multi-modal MRI images provide multi-dimensional information for the diagnosis and treatment of diseases.However,it takes a lot of economic and time costs to obtain multi-modal images of the same subject at the same time,which leads to too long scanning time and increases the difficulty of obtaining high-quality MRI images in the abdomen with complex motion.So if the MRI with short scan time can be used to synthesize other modal images,the scanning time and motion artifacts can be reduced,and high quality multi-modal images can be obtained at the same time,which is of great significance in abdominal MRI imaging.Therefore,based on deep learning algorithm,this paper proposes a multi-modal image synthesis network for abdominal T2WI: Dual channel network based on residual divided attention(DRSA-Unet),which synthesizes T2 WI through abdominal T1 WI and down-sampled T2 WI.The scan time of T2 WI was reduced to one quarter of that on real T2 WI.he main work is as follows:1.The preprocessing of data and the pre-training of network for the design of abdominal MRI multi-modal image synthesis network.The multi-modal images in CHAOs abdominal data set used in this paper have some problems,such as small data volume and without registration.Therefore,this paper firstly normalized the data,and then proposed an inter-layer registration algorithm based on liver segmentation to achieve inter-layer matching of multi-modal images.Finally,the data set was expanded through data enhancement.In addition,based on the principle of MRI imaging and K-space image reconstruction technology,the method of T2WI down-sampling is introduced.2.Introduces the medical image synthesis algorithm based on deep learning,and selects the excellent performance of U-Net and Dense-Unet network for improvement in this paper.Using the improved U-Net network,the data input method of cross-modal synthesis abdominal T2 WI and the effect of inter-layer registration were tested and verified by pre-experiment.3.Design the abdominal MRI multimodal image synthesis network: DRSA-Unet,based on deep learning.The segmentation attention residuals module and compression excitation attention module were introduced into the network,which expanded the receptive field to learn multi-scale information,and improved the ability of the network to extract,compress and integrate multi-channel information in the multi-modal image synthesis task.The experimental results show that the performance and image quality of DRSA-Unet network are better than that of the U-Net and Dense-Unet in synthesizing T2WI using T1WI＿Out and down-sampled T2WI.