Comparing the Training Performance of a Deep Neural Network for Accelerated MRI Reconstruction Using Synthesized and Realistic k-Space Data

Magnetic Resonance Imaging (MRI) is a powerful medical imaging modality used as a diagnostic tool. There is a steady rise in the imagining examination. Trends from 2000 - 2016 showed that nearly 16 million to 21 million patients had enrolled annually in various US health care systems. The number of MRIs per 1000 increased from 62 per 1000 to 139 per 1000 patients from 2000 to 2016. MR images are usually stored in Picture Archiving and Communication Systems (PACS) in Digital Imaging and Communication in Medicine (DICOM). DICOM format includes a header and imaging data. MRI k-space is the raw data obtained during the MR signal acquisition. The file size of complex MR data is huge. It is generally transformed into the anatomical imaging data, and raw data is discarded and not transferred to the PACS. The abundant DICOM data has the potential to be used for training neural networks. Deep Neural Network models depend on the extensive training datasets. DICOM images are magnitude images without the image phase. It is essential to understand the effect of missing image phase information to use the DICOM data for this training task effectively. My thesis attempts to compare a deep neural network's performance for accelerated MRI reconstruction using the k-space to DICOM only data. MR imaging offers a great deal of control to the user to acquire the data and reconstruct the clinical images. All this comes at the cost of an increase in the acquisition time. Typical scan times are between 30 to 40 mins. Scan times go up to 60 mins if a contrast agent needs to be administered. Such long acquisition times are not only expensive but a cause of inconvenience to the subject as it is impossible to stay motionless in the bore during the whole duration. Two areas are of interest to reduce the scan time, (i) accelerated acquisition and (ii) fast and efficient reconstruction. Methods like compressed sensing and parallel imaging are used to accelerate MRI acquisition. Compressed sensing achieves scan acceleration by overcoming the requirement of Nyquist sampling criteria. An undersampling pattern like the Poisson Disk undersampling pattern is used to acquire an incoherent random sparse signal instead of the full k-space. The "sigpy.mri" python library's "Poisson" API was used to simulate this undersampling. This Python API generates a variable-density Poisson-disc sampling pattern. Compressed Sensing theory mentions that image reconstruction would be possible using signals less than the number indicated by Nyquist as long as the k-space undersampling is done incoherently, which does not lead to structural aliasing when the anatomical image is constructed. This algorithm combines the undersampling with partial Fourier imaging. This API uses a fully sampled calibration region at the center of the k-space in addition to the acceleration factor. The acceleration factor is used for undersampling the region outside the fully sampled center region. Poisson disk undersampling does random sampling while constraining the maximum and minimum distance. This scheme leads to incoherent sampling and avoids structural artifacts. After the image acquisition comes, the reconstruction of the fully sampled k-space or the anatomical image with good SNR. A deep-learning neural network was trained to perform the reconstruction of the retrospectively undersampled data. The undersampled raw k-space data's training performance is compared with that of the undersampled k-space data obtained from the DICOM data. Our experiments have shown that the magnitude obtained from raw k-space data has consistently shown better initial training performance and faster convergence when compared to the magnitude image obtained from the DICOM image. It is also observed that after training enough epochs, the performance of the model trained using raw data is comparable to that of the DICOM images. The significance of this finding is in the fact that the abundantly available DICOM data can be used to train a deep neural network to perform reconstruction of the undersampled k-space. FastMRI is a research project from Facebook AI(FAIR) and NYU Langone Health. The dataset for this project is publicly available. This dataset has two types of scans, knee MRI and brain MRI. For this work, we have used single coil knee MRI data. For performing the training, 2D slices from these images are used from the training dataset's single-coil knee MRI volumes. The training dataset has 973 volumes and a total of 34,742 slices