Unsupervised Generative Adversarial Networks For CT/MRI Conversion Based On Adaptive And High-Low Frequency Convolution

Radiotherapy is currently one of the effective means of treating malignant tumors,and physicians need integrated information from different modality medical images(e.g.,CT images,MRI images,etc.)to make an accurate diagnosis.Therefore,a framework that can interchange between CT images and MRI images is of great significance.For physicians,it can improve the efficiency of diagnosis and also reduce the complexity and duplication of work to a certain extent;for patients,the detection process becomes more convenient.Currently,the traditional supervised CT/MRI image translation framework relies on medically aligned datasets.However,it is difficult to obtain in real medical scenarios.In addition,the images translated by unsupervised image translation frameworks often suffer from blurring and loss of key details.To address the above problems and based on the feature that the main feature information of CT/MRI images is concentrated in high frequency information,this thesis proposes an unsupervised generative adversarial network model based on high-and low-frequency convolution—Cycle-SOAGAN.The main contribution points are as follows.(1)Using high-and low-frequency group convolution.The high-frequency information helps to maintain the original feature information of the image,and the low-frequency information can expand the perceptual field of view of the convolutional layer to obtain more contextual information.(2)Use self-attentive mechanism.It effectively captures the remote dependencies between features and achieves the purpose of focusing on important regions for intensive transformation.(3)Use adaptive layer/instance normalization.Adaptively balance the ratio of layer normalization and instance normalization to guide the network to flexibly control the shape and texture changes.(4)A new dataset of CT/MRI datasets was compiled and produced based on data provided by hospitals and CT/MRI brainscans,and qualitative and quantitative analyses of this method were performed in the new dataset.In this thesis,comparison experiments and ablation experiments were conducted with the current unsupervised translation framework,and the effectiveness of this method was proved by the evaluation index and visualization results,and the purpose of two-way high-quality translation of CT images and MRI images was achieved.