Dual-domain interaction Transformer for MRI reconstruction
Partial k-space sampling is a primary method for accelerating Magnetic Resonance Imaging(MRI).Reconstruction of high-quality MRI images from undersampled data has important applications in clinical diagnosis and research analysis.In recent years,deep learning-based approachs have made some prog-ress in MRI reconstruction.However,the networks that solely focus on the image domain or the frequency domain cannot effectively utilize the features of both domains to improve reconstruction quality.Furthermore,existing dual-domain reconstruction methods often lack interaction and fusion between the domains,limiting their reconstruction performance.To address these issues,a dual-domain interaction Transformer network is proposed in this paper for MRI reconstruction.The proposed method extracts dual-domain features with Transformer and utilizes cross-domain attention to guide the features fusion,enabling efficient extraction and integration of the complementary information from both domains.Specifically,since each point in the fre-quency domain data corresponds to all pixels in image domain,the 1×1 convolution in k-space is utilized to extract global features,and while a window-based Transformer is designed to model local features in the im-age domain.Then a fusion module based on cross attention is introduced to guide the fusion of dual-domain features and integrate cross-domain complementary information.Experimental results demonstrate that the method in this paper outperforms other baseline reconstruction methods on publicly available datasets.More-over,ablation studies validate the effectiveness of the proposed network modules.
国家科技期刊平台
NSTL
万方数据
