Research map / Image generation / Autoregressive and latent generation
Adaptive Sampling of k-Space in Magnetic Resonance for Rapid Pathology Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ASMR, an RL (PPO) agent that adaptively selects k-space columns to maximize pathology classification from sparse MR k-space, within ~2% of a fully sampled image classifier using ~8% of k-space.
Problem
MR scans are slow, and reconstruction-based or heuristic/non-adaptive k-space sampling patterns are suboptimal for detecting pathology from under-sampled k-space at population-scale screening.
Approach
Frame sequential Cartesian k-space column selection as an MDP and train a PPO policy whose reward is the log-likelihood of the label under a frozen k-space classifier (KSpace-Net), with dynamic action masking and label-balanced environment sampling.
Key results
- ASMR is within 2% of a fully sampled image-based classifier on 6 of 8 pathology tasks while using about 8% of k-space (12.5x acceleration).
- ASMR outperforms EMRT on 6 of 8 tasks and LOUPE/DPS on most tasks, with at least 2.5% AUC gain on 7 of 8 tasks versus learned probabilistic methods.
- ASMR beats reconstruction-optimized policies by about 7% average AUROC on brain pathologies.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use adaptive classification-optimized sampling instead of learned non-adaptive sampling | ASMR versus DPS, LOUPE, and EMRT · Knee and Brain pathology tasks | AUC: atleast 2.5% | Helped |
| Use adaptive classification-optimized sampling instead of learned non-adaptive sampling | ASMR versus DPS, LOUPE, and EMRT · Prostate CS-PCA | AUC | Hurt |
| Optimize the sampling policy for classification rather than image reconstruction | ASMR versus reconstruction-optimized policy · Brain pathology tasks | AUC: 7% | Helped |
| Use ASMR rather than a non-adaptive greedy sequence | Sequential k-space sampling · Knee, Brain, and Prostate datasets | AUC: 1.87%, 7.01%, and 9.82% | Helped |
| Use ASMR rather than variable-density random sampling | Sequential k-space sampling · Knee, Brain, and Prostate datasets | AUC: 1.33%, 2%, and 2.25% | Helped |
| Use ASMR instead of a fully sampled image-based classifier | Pathology prediction · 6 out of 8 classification tasks | AUC: within $2\%$ | Mixed |
| Add label-balanced sampling of training environments (ablation) | ASMR · Knee dataset across evaluated sampling rates | AUC | Helped |
| Use explicit action-space masking instead of penalizing repeated measurement selections (ablation) | ASMR · Knee dataset | AUC | Helped |
Surprising findings
- Reconstruction-optimized sampling policies often select masks that are less useful for pathology classification than policies trained on the classification reward.
- Without label-balanced environment sampling, the learned policy becomes biased toward normal patients and performance drops across sampling rates.
Limitations
- Uses only single-coil k-space (emulated single-coil for brain and prostate), not multi-coil data.
- Slice-level classifiers only; no volumetric classification or segmentation.
- Retrospective under-sampling with limited dataset size.
Experimental setup
Datasets: FastMRI knee FastMRI brain Prostate MR (CS-PCA)
Benchmarks: FastMRI knee pathology labels FastMRI brain pathology labels Prostate CS-PCA
Metrics: AUROC AUC balanced accuracy sensitivity specificity NPV
Baselines: EMRT LOUPE DPS VDS random policy non-adaptive greedy sequence reconstruction-optimized greedy policy fully-sampled image-based ResNet-50
Builds on: PPO EMRT KSpace-Net LOUPE DPS
Closest papers by meaning
- AeSPa : Attention-guided Self-supervised Parallel Imaging for MRI Reconstruction (CVPR 2025 · similarity 0.86)
- Autoregressive Image Diffusion: Generation of Image Sequence and Application in MRI (NeurIPS 2024 · similarity 0.86)
- Motion Artifact Removal in Pixel-Frequency Domain via Alternate Masks and Diffusion Model (AAAI 2025 · similarity 0.83)
- Fast MRI for All: Bridging Access Gaps by Training without Raw Data (NeurIPS 2025 · similarity 0.83)
- Decomposed Diffusion Sampler for Accelerating Large-Scale Inverse Problems (ICLR 2024 · similarity 0.83)
- D2SA: Dual-Stage Distribution and Slice Adaptation for Efficient Test-Time Adaptation in MRI Reconstruction (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.