Frequency-Hierarchical Active k-Space Sampling for Diagnostic MRI

2026-07-23 · Source: cs.CV updates on arXiv.org · Field: Science & Research — Health & Medical Research, Mathematics & Computational Sciences, Medical Imaging AI · Depth: Expert, long

Summary

HieraSample is a novel task-driven active k-space sampling framework designed to accelerate diagnostic MRI without compromising accuracy. It addresses the critical challenge of efficiently distributing a limited sampling budget across spatial frequencies, which carry distinct anatomical and pathological information. The framework employs a cosine-annealed curriculum, progressively reducing the acceleration factor from 20× to 4× over 80 acquisition steps, while consistently preserving a fully-sampled low-frequency disk. A Mamba-based policy, informed by dual disease and severity classifiers, then strategically selects individual high-frequency coordinates. This policy is driven by a reward system that quantifies the per-sample reduction in class-weighted cross-entropy, directly linking sampling decisions to improved diagnostic confidence. Evaluated on the fastMRI+ knee benchmark, HieraSample achieved diagnostic accuracy matching the fully-sampled oracle for ACL diagnosis across 4× to 10× acceleration. Furthermore, it demonstrated a significant improvement of up to 20.4 AUC points on ACL severity compared to the ASSDM baseline.

Key takeaway

For machine learning engineers developing accelerated MRI techniques, HieraSample demonstrates a superior approach to k-space sampling. You should consider hierarchical sampling strategies that prioritize low-frequency data while actively selecting high-frequency points based on diagnostic task relevance. This method significantly improves diagnostic accuracy for conditions like ACL severity, outperforming uniform sampling. Implement Mamba-based architectures for efficient, long-sequence policy rollouts in your active acquisition systems.

Key insights

Active MRI sampling can optimize diagnostic accuracy by hierarchically prioritizing k-space frequencies and using Mamba-based policies.

Principles

• Low-frequency k-space data provides essential anatomical context.
• High-frequency k-space data drives fine pathological assessment.
• Reward functions should directly reflect diagnostic improvement.

Method

HieraSample uses a cosine-annealed curriculum to reduce acceleration from 20× to 4× over 80 steps, maintaining a low-frequency disk. A Mamba policy then selects high-frequency coordinates based on disease/severity classifier features, optimizing for cross-entropy reduction.

In practice

• Preserve low-frequency k-space for anatomical context.
• Use Mamba blocks for efficient sequential decision-making.
• Target diagnostic accuracy, not just image quality metrics.

Topics

• Active MRI Sampling
• k-Space Undersampling
• Mamba State-Space Models
• Diagnostic Imaging
• fastMRI+ Benchmark
• Reinforcement Learning

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Frequency-Hierarchical Active k-Space Sampling for Diagnostic MRI — HieraSample optimizes MRI k-space sampling by prioritizing low-frequency data and actively selecting high-frequency details for improved diagnostic accuracy.
• PhenSPINE: A Standardized Benchmark for Spine Pathology Diagnosis — Positional Encoding and Sagittal T2-weighted MRI are crucial for accurate deep learning-based spinal pathology diagnosis.
• Dual-Prior Guided Null-Space Learning with Mixture-of-Splines for Arbitrary Medical Slice Super-Resolution — DP-NSL reformulates super-resolution as a constrained recovery, using dual priors for measurement consistency and geometric continuity.
• MambaBack: Bridging Local Features and Global Contexts in Whole Slide Image Analysis — MambaBack is a hybrid WSI analysis model that combines Mamba and Gated CNNs for efficient, accurate, and memory-optimized cancer diagnosis.
• Vision-Language Model-Guided Deep Unrolling Enables Personalized, Fast MRI — PASS uses a VLM to guide personalized MRI sampling and reconstruction for faster, anomaly-aware imaging.
• MRecover: A Conditional Generative Model for Recovering Motion-Corrupted MR images Using AI Generated Contrast — MRecover uses a conditional generative model to reconstruct motion-corrupted MRI, significantly increasing analyzable data and diagnostic power.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.