Enhanced Deep-Learning-Based Magnetic Resonance Image Reconstruction by Leveraging Prior Subject-Specific Brain Imaging: Proof-of-Concept Using a Cohort of Presumed Normal Subjects

Deep learning models have shown potential for reconstructing undersampled, multi-channel magnetic resonance (MR) image acquisitions. Recently proposed methods, however, have not leveraged information from prior subject-specific MR imaging sessions. Such data are often readily available through a picture archiving and communication system (PACS). We propose a flexible three-step method to incorporate this prior information into an enhanced deep-learning-based reconstruction process. The method consists of Step 1: an initial reconstruction; Step 2: registration of the previous scan to the initial reconstruction; and Step 3: an enhancement network. Training and testing used longitudinally acquired, three-dimensional, T1-weighted brain images acquired with different acquisition parameters. We tested our networks using data from 2808 images (obtained in 18 subjects) under four different acceleration factors (R = {5, 10, 15, 20}). Our enhanced reconstruction (Steps 1-3) produced higher-quality images: structural similarity and peak signal-to-noise ratio increased, and normalized root mean squared error decreased on average by 16.5%, 7.0% and 21.1%, respectively, compared to the non-enhanced reconstruction (Step 1 only) under the same network capacity as the enhanced reconstruction model. These differences were statistically significant (p < 0.001, Wilcoxon signed-rank test). Further volumetric analysis performed on key brain regions (brain, white matter, gray matter and cortex) indicated that our enhanced images had better volume agreement with the fully sampled reference images compared to the non-enhanced images. Our enhanced images for R = 20 were comparable to the non-enhanced images for R = 10 demonstrating that our proposed method can use prior scan information to further accelerate MR examinations.