Slice-Aware Sampling in CS-based Deep Learning Brain MRI Reconstruction

Deep Neural-based reconstruction for Compressed Sensing (CS) has emerged as a powerful approach for accelerating brain Magnetic Resonance Imaging (MRI), minimizing scan duration and patient discomfort, surpassing traditional methods. Among the numerous available architectures, ADMM-CSNet stands out as one of the most effective state-of-the-art solutions. However, this approach is affected by a high standard deviation in the reconstruction accuracy. In this work, we investigate the cause of this variability. Specifically, we show that brain MRI itself has a high variability in information content, meaning that the first and last sections are relatively simple to reconstruct, while central views are more complex.To overcome this problem, we adopt the Weighted Random Sampling (WRS) strategy, enhancing both reconstruction quality and robustness. Although this is not typically applied to this kind of dataset, it allows a refinement of the ADMM-CSNet framework, directing the network’s learning process to prioritize central slices over lateral slices. Experimental results demonstrate that the proposed approach consistently outperforms the baseline ADMM-CSNet and other state-of-the-art methods in terms of mean achieving a 22% lower standard deviation than the original model. Finally, the experimental results on pathological MRI images further validate the efficacy of the proposed method, showing an average PSNR gain of 3 dB over the original baseline.