GAMBAS: Generalised-Hilbert Mamba for Super-resolution of Paediatric Ultra-Low-Field MRI

Levente Baljer, Ula Briski, Robert Leech, Niall J Bourke, Kirsten A Donald, Layla E Bradford, Simone R Williams, Sadia Parkar, Sidra Kaleem, Salman Osmani, Sean CL Deoni, Steven CR Williams, Rosalyn J Moran, Emma C. Robinson, František Váša

Keywords: Mamba, Hilbert curve, MRI, Deep Learning, Neurodevelopment

Abstract: Magnetic resonance imaging (MRI) is critical for neurodevelopmental research, however access to high-field (HF) systems in low- and middle-income countries is severely hindered by their cost. Ultra-low-field (ULF) systems mitigate such issues of access inequality, however their diminished signal-to-noise ratio limits their applicability for research and clinical use. Deep-learning approaches can enhance the quality of scans acquired at lower field strengths at no additional cost. For example, Convolutional neural networks (CNNs) fused with transformer modules have demonstrated a remarkable ability to capture both local information and long-range context. Unfortunately, the quadratic complexity of transformers leads to an undesirable trade-off between long-range sensitivity and local precision. We propose a hybrid CNN and state-space model (SSM) architecture featuring a novel 3D to 1D serialisation (GAMBAS), which learns long-range context without sacrificing spatial precision. We exhibit improved performance compared to other state-of-the- art medical image-to-image translation models. Our code is made publicly available at https://github.com/levente-1/GAMBAS.

Primary Subject Area: Application: Neuroimaging

Secondary Subject Area: Image Synthesis

Paper Type: Both

Registration Requirement: Yes

Submission Number: 135