Browsing by Author "Dubljevic, Natalia"

Now showing 1 - 1 of 1
  • Thumbnail Image
    ItemOpen Access
    Assessing Multi-Coil Deep-learning-based MR Image Reconstruction
    (2025-01-11) Dubljevic, Natalia; Frayne, Richard; Souza, Roberto; Bayat, Sayeh; Levan, Pierre
    Magnetic resonance (MR) imaging is a valuable non-invasive imaging technology essential for diagnostics and patient health management. However, scan times are often quite long, affecting both patient throughput and comfort. To help accelerate the process, parallel imaging uses an array of receiver coils to reduce the amount of data acquired. Recently deep learning (DL)-based MR image reconstruction has become popular as it allows for further scan acceleration. With its advent, new questions are raised regarding optimal coil design and processing strategies. This work investigates whether traditional geometric coil constraints imposed by parallel imaging can be relaxed when using DL-based reconstruction methods as compared to non-DL approaches. We also explore whether to first combine channels (coil-combined approach) to enhance model generalizability or to keep channel processing separate (all-coil approach) to fully utilize multi-channel information. Two sets of head coil profiles (8-channel and 32-channel geometries) were evaluated across three methods: a DL model, conjugate gradient SENSE (CG-SENSE), and L1-wavelet compressed sensing (CS). These methods were compared using both quantitative metrics and visual assessment as coil overlap increased. Results showed that as coil overlap increased, performance significantly decreased (p<0.001) across all methods, with the largest impact seen in CG-SENSE. DL-based reconstructions consistently outperformed their non-DL counterparts, with minimal changes in performance across coil overlap and signal-to-noise ratio (SNR). While CS demonstrated better robustness to coil overlap than CG-SENSE, it produced inferior reconstructions characterized by blurriness. Additionally, we assessed three popular DL architectures using both coil-combined and all-coil strategies on brain MR images in and out of distribution. All-coil styles slightly improved in-distribution performance, such as when reconstructing data from healthy individuals, while coil-combined designs better generalized to unseen pathological data. These findings suggest that DL-based reconstruction can produce high-quality images that are robust to coil overlap, offering the potential to relax geometric coil design constraints. Additionally, a coil-combined processing approach may be preferable when considering clinical use.