Application Of Convolutional Neural Network In Magnetic Resonance Imaging Reconstruction And Quality Evaluation

In recent years,deep learning has achieved great success in computer vision field,especially when applying Convolutional Neural Network(CNN)to image recognition and detection.CNN is composed of multiple convolutional layers,which can automatically learn different feature maps from a large number of data,capture the nonlinear mapping between the input and the output and no longer rely on manual feature extraction.Inspired by the achievements of deep learning,this paper tried to use CNN to solve several problems in MRI:1.Long scan time is a major weakness of MRI.To shorten the scan time,various undersampling schemes have been proposed,each of which will inevitably introduce different artifacts into the images.For example,a truncated k-space will lead to Gibbs ringing artifacts.In this work,CNN was used to process the undersampled MRI images to eliminate artifacts while keeping the structural details in the images.2.Quantitative susceptibility map(QSM)is a new technique for quantitatively measuring the magnetic susceptibility of tissues.Unfortunately,calculating the distribution of magnetic susceptibility from phase map is an ill-posed inverse problem and the reconstructed images are always accompanied with serious streaking artifacts.In this paper,CNN was trained with a bunch of image pairs with and without artifacts and was used to suppress artifacts and reconstruct high quality QSM.3.Motion in the MRI scanning process will result in artifacts such as image distortion,blurring or absence,which makes the image unqualified for clinical diagnosis.This problem is especially severe in hospitals with a high patient flow,where enough instruction and training of the patient often lacks.To make it worse,the issue would not be found out till a radiologist started to read the images,when the scan had finished for some time.Thus,it will be very useful if we can evaluate the image quality while the scan is in process and remind the operator of the possible quality issues in real time.To achieve this,an CNN was trained to extract region of interest(ROI)from the images and a subsequent CNN model was trained to evaluate the image quality from the image blocks in the ROI.From the results of this work,we can see that CNN exhibits its unique advantage over the traditional image processing algorithms in both MRI image reconstruction and quality evaluation.CNN shows great potential in applications in MRI.