Research On CS-MRI Reconstruction Methods For Regions Of Interest

Magnetic resonance imaging for specific regions of interest has important value and significance for people’s diagnosis,treatment and prognosis,and is one of the important tasks in the intelligent analysis of medical images.MRI is limited by its slow imaging speed,and compressed sensing technology can reconstruct a clear image from a small amount of K-space acquisition data to reduce the data acquisition time of MRI.Therefore,CS-MRI reconstruction has attracted the attention and research of many scholars.In most actual medical imaging application scenarios,doctors often only focus on specific areas such as human tissues or lesions in the images,and these specific areas are rich in auxiliary diagnostic information.However,the existing compressed sensing magnetic resonance imaging reconstruction method does not make full use of sufficient information of the region of interest in the two stages of data acquisition and image reconstruction,and cannot accurately reconstruct the region of interest.With limited K-space sampling resources,in order to effectively utilize the imaging information associated with this task to improve the quality of fast MRI in specific regions,this paper proposes a deep convolutional neural network framework that combines sampling mode and reconstruction process to optimize.Through a learnable partial K-space sampler,the framework can obtain K-space adaptive sampling templates for specific regions.In the reconstruction phase,the framework uses a cascaded convolutional neural network as the reconstructor,and intersperses a data fidelity layer to correct errors.In order to make full use of the location information of the region of interest,this paper proposes a reconstruction model based on the attention mechanism and a reconstruction model based on multi-task learning.The reconstruction model based on the attention mechanism uses the pre-segmentation model to obtain the location information of the region of interest,and introduces the attention module of feature calibration.The reconstruction model based on multi-task learning adds a position prediction branch to introduce high-level semantic information to guide the accurate reconstruction of the region of interest.The main innovations of this article include:Firstly,the proposed framework is the first deep learning model with joint optimization of K-space sampling and image reconstruction for regions of interest.Secondly,we explore two ways to introduce information from regions of interest:attention mechanism and multi-task learning.Thirdly,we conduct experimental verification on the abdominal organ segmentation dataset and the brain tumor segmentation dataset.Compared with other general and fast MRI reconstruction methods for regions of interest,our model achieved the most advanced performance.In the case demonstration,our method can recover some small details of the region of interest at a low sampling rate.