»ã±¨±êÌâ (Title)£ºGeneralizable and interpretable MRI reconstruction with high data heterogeneity£¨ÓµÓиßÊý¾ÝÒìÖÊÐԵķº»¯ºÍ¿ÉÚ¹ÊÍÐÔMRI³Á½¨£©

»ã±¨ÈË (Speaker)£º³ÂÔÏ÷£¨ÃÀ¹ú·ðÂÞÀï´ï´óѧƽÉú½ÌÊÚ£©

»ã±¨¹¦·ò (Time)£º2024Äê11ÔÂ3ÈÕ(ÖÜÈÕ) 10:00-12:00

»ã±¨µØÖ· (Place)£ºÐ£±¾²¿ FÂ¥ËÄÂ¥Êýѧϵ»áÉÌÊÒ

Ô¼ÇëÈË(Inviter)£ºÅíÑÇР½ÌÊÚ

Ö÷°ì²¿ÃÅ£ºÀíѧԺÊýѧϵ

»ã±¨ÌáÒª£ºDeep learning methods have demonstrated promising performance in a variety of image reconstruction problems. However, task specific and extremely data demanding are still a major challenging in practical applications. In this work we introduce a generalizable MRI reconstruction method with diverse dataset to tackle those problems. Our approach proposes a variational model, in which the learnable regularization function is parameterized by two sets of parameters: a task-invariant set for common feature encoding and a task-specific part to account for the variations in the heterogeneous data. Then, we generate a neural network, whose architecture follows exactly a convergent learned optimization algorithm for solving the nonconvex and nonsmooth variational model. The network is trained by a bilevel optimization algorithm to prevent overfitting and improve generalizability. A series of experimental results on heterogeneous MRI data sets indicate that the proposed method generalizes well to the reconstruction problems whose undersampling patterns and trajectories are not present during training.