Arbitrary-Scale Anisotropic 3D MRI Super-Resolution with Gaussian Prototypes

Key points

• Reducing MRI acquisition time improves patient comfort, reduces motion artefacts, and increases scanner throughput.1
• It combines a learned dictionary of anisotropic 3D Gaussian prototypes with scale-conditioned frequency encoding and an exact measurement-consistency projection, keeping reconstructions consistent with the acquired low-resolution data.1
• We evaluate GSMR3D against geometry-specific specialists for isotropic, in-plane, and through-plane super-resolution using reconstruction quality and downstream tasks: brain tumour segmentation, brain-metastasis detection, and brain parcellation.1
• GSMR3D offers a flexible approach to MRI scan-time reduction with one model per contrast across geometries while preserving information relevant to downstream tasks.1

Sources (1)

• [1]Arbitrary-Scale Anisotropic 3D MRI Super-Resolution with Gaussian Prototypes
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 17, 08:18 PM

  We present GSMR3D, a continuous volumetric model that reconstructs arbitrary isotropic and anisotropic resolutions with a single checkpoint per MRI contrast.

  Reducing MRI acquisition time improves patient comfort, reduces motion artefacts, and increases scanner throughput.

Extractive summary: sentences quoted from the sources.