Reduction in scan time is a major direction of development in magnetic resonance imaging (MRI) to decrease patient discomfort, minimize motion-related artifacts and improve patient throughput. Nowadays, it has become common practice to undersample k-space in order to reduce scan time by decreasing phase-encoding steps. Parallel imaging (PI), one of the widely accessible techniques, employs alternate omission of phase-encoding steps and reconstructs images from undersampled k-space using coil-sensitivity or auto-calibrated methods [1,2]. However, residual aliasing and noise enhancement are trade-offs for high acceleration factors in PI [3]. Compressed sensing (CS) is another recently implemented scheme in clinical practice, acquiring incoherently undersampled data and iteratively reconstructing images with sparsity transformation for de-noising and data consistency term for data fidelity [4,5]. However, CS is less effective for 2-dimensional (2D) fast spin-echo (FSE) images due to low sparsity and randomness of 2D Cartesian sampling itself [6]. In addition, long reconstruction time and blurring or ringing artifacts generation are difficult to address for its clinical use [7,8].

Recently, deep learning (DL)-based MR reconstruction showed promising results in accelerated imaging [[9], [10], [11]]. DLs are typically trained in a supervised manner, considering data consistency based on a PI model as well as regularizations to reduce undersampling-related noise and artifacts. The regularizations are therefore adapted to provide the best performance on the training dataset, in contrast to CS which chooses regularizations based on priori assumptions on sparsity of the reconstructed images [4,12]. The obtained weights of DL networks can be exported and prospectively applied in clinical practice under the assumption of the representative training dataset and generalization of trained reconstruction [9,13].

According to previous studies, DL-based reconstruction has potential to significantly reduce scan time while preserving the high image quality in spine or knee [14,15]. However, this new technique needs more validation for optimal parameter setting and diagnostic equivalence in universal disease before clinical deployment. Therefore, the purpose of our study was to compare the image quality and to assess the interchangeability and agreement for degenerative lesion detection between conventional FSE and DL-reconstructed T2-weighted image (T2-WI) in cervical (C-) and lumbar (L-) spine imaging.