Deep subspace learning for efficient reconstruction of spatiotemporal imaging data

Christopher Michael Sandino, Frank Ong, Siddharth Srinivasan Iyer, Adam Bush, Shreyas Vasanawala

Keywords: Spatiotemporal Reconstruction, Magnetic Resonance Imaging, Memory-efficient

TL;DR: A novel unrolled neural network architecture is proposed which reduces memory footprint by reconstructing compressed representations of high-resolution spatiotemporal imaging data instead of the data directly.

Abstract: Model-based deep learning approaches, such as unrolled neural networks, have been shown to be effective tools for efficiently solving inverse problems. However, the memory costs of training unrolled networks remain high, especially when the target data is high-resolution and high-dimensional. This often requires trade-offs in either network depth to reduce model size, or data resolution to reduce data size. To address this, we propose DL-Subspace - a novel unrolled network architecture which reduces memory usage by solving for a compact, low-dimensional representation of the target instead of the target itself. DL-Subspace is applied to accelerated magnetic resonance image reconstruction, demonstrating up to 4$\times$ higher memory efficiency and 4$\times$ faster inference speed while maintaining similar image quality as conventional unrolled networks.