Adaptive Sampling of k-Space in Magnetic Resonance for Rapid Pathology Prediction

TL;DR

This work tackles the challenge of long MR scan times by reframing data collection as a sequential decision process that directly optimizes pathology prediction from sparse k-space data. ASMR learns a reinforcement learning policy, via Proximal Policy Optimization, to sequentially select k-space columns with a reward based on the log-likelihood of pathology labels given the undersampled data, bypassing image reconstruction. Across knee, brain, and prostate datasets, ASMR achieves near-fully sampled performance at only $8\%$ of the k-space and outperforms state-of-the-art non-adaptive and reconstruction-optimized sampling methods on most tasks. The results demonstrate the potential for reconstruction-free, adaptive acquisition to enable rapid MR-based screening at population scale, while highlighting avenues for extending to multi-coil data and volumetric/pathology segmentation applications.

Abstract

Magnetic Resonance (MR) imaging, despite its proven diagnostic utility, remains an inaccessible imaging modality for disease surveillance at the population level. A major factor rendering MR inaccessible is lengthy scan times. An MR scanner collects measurements associated with the underlying anatomy in the Fourier space, also known as the k-space. Creating a high-fidelity image requires collecting large quantities of such measurements, increasing the scan time. Traditionally to accelerate an MR scan, image reconstruction from under-sampled k-space data is the method of choice. However, recent works show the feasibility of bypassing image reconstruction and directly learning to detect disease directly from a sparser learned subset of the k-space measurements. In this work, we propose Adaptive Sampling for MR (ASMR), a sampling method that learns an adaptive policy to sequentially select k-space samples to optimize for target disease detection. On 6 out of 8 pathology classification tasks spanning the Knee, Brain, and Prostate MR scans, ASMR reaches within 2% of the performance of a fully sampled classifier while using only 8% of the k-space, as well as outperforming prior state-of-the-art work in k-space sampling such as EMRT, LOUPE, and DPS.

Figures (11)

• Figure 1: Overview of ASMR. Given a set of prior samples from the MR scanner, ASMR proposes the next sample to collect.These sampling steps are repeated for a fixed set of iterations $T$, at which point the collected samples are used by a classifier to predict the presence of a fixed set of pathologies. ASMR is trained to optimize classification performance and sidesteps reconstruction altogether.
• Figure 2: Cartesian Sampling on k-space at different sampling rates. The leftmost image shows the complete k-space followed 2.5%, 5% and 10% sampling rates (from left to right)
• Figure 3: ASMR Training.ASMR takes an initial sub-sampled k-space $\mathbf{s}_t$ as input and proposes the next sample $\mathbf{a}_t$ to generate the next state $\mathbf{s}_{t+1}$. This state $\mathbf{s}_{t+1}$ is used by the reward model to compute the log-likelihood $q_\phi(\mathbf{y} \mid \mathbf{x}_{\mathbf{s}_{t+1}})$, the reward for the actor. We repeat these steps for a fixed number of iterations $T$.
• Figure 4: AUROCs obtained by ASMR compared to Learned Non-Adaptive Methods; the horizontal dotted line in red denotes the performance of an image-based classifier (which uses the entire k-space data). ASMR outperforms LOUPE and DPS for 6 out of 8 tasks, and outperforms EMRT on 7 out of 8 tasks. All results are computed over 5 seeds, and plotted with their means and standard deviations.
• Figure 5: AUROCs obtained by ASMR compared to sequential sampling methods such as VDS random and non-adaptive greedy sequence. In the low sampling regime, ASMR consistently outperforms both baselines, while either outperforming or matching their performance at higher sampling rates. All results are computed over 5 seeds, and plotted with their means and standard deviations.