Active Sampling for MRI-based Sequential Decision Making

TL;DR

Active Sampling for MRI-based Sequential Decision Making addresses reducing MRI acquisition time by actively sampling undersampled k-space using a multi-objective reinforcement learning framework. The method models sequential clinical decisions—disease presence and severity—as a Partially Observable Markov Decision Process and optimizes a step-wise lexicographic reward to share informative k-space lines across tasks. Experiments on ACL sprain detection and cartilage thickness loss demonstrate that the approach achieves competitive diagnostic performance with fully sampled data while saving substantial k-space samples, and performs sequential diagnosis within a single scanning session. The work advances direct, sample-efficient MRI inference and offers open-source code for reproducibility.

Abstract

Despite the superior diagnostic capability of Magnetic Resonance Imaging (MRI), its use as a Point-of-Care (PoC) device remains limited by high cost and complexity. To enable such a future by reducing the magnetic field strength, one key approach will be to improve sampling strategies. Previous work has shown that it is possible to make diagnostic decisions directly from k-space with fewer samples. Such work shows that single diagnostic decisions can be made, but if we aspire to see MRI as a true PoC, multiple and sequential decisions are necessary while minimizing the number of samples acquired. We present a novel multi-objective reinforcement learning framework enabling comprehensive, sequential, diagnostic evaluation from undersampled k-space data. Our approach during inference actively adapts to sequential decisions to optimally sample. To achieve this, we introduce a training methodology that identifies the samples that contribute the best to each diagnostic objective using a step-wise weighting reward function. We evaluate our approach in two sequential knee pathology assessment tasks: ACL sprain detection and cartilage thickness loss assessment. Our framework achieves diagnostic performance competitive with various policy-based benchmarks on disease detection, severity quantification, and overall sequential diagnosis, while substantially saving k-space samples. Our approach paves the way for the future of MRI as a comprehensive and affordable PoC device. Our code is publicly available at https://github.com/vios-s/MRI_Sequential_Active_Sampling

Figures (6)

• Figure 1: Illustration of our proposed sequential diagnostic framework for active MR acquisition and its motivating evidence: in preliminary experiments, we found similarities between the sampling trajectories of disease detection and severity quantification. We trained policies du2024mri optimized independently for each task and obtained sampling trajectories shown in (a), and then measured their similarity in (b). We observe an increasing Pearson correlation between sampling steps $4$ and $14$, indicating that in this range $k$-space samples contribute simultaneously to both tasks. These observations motivate our approach (c) to learn a sampling policy that jointly optimizes two tasks, taking into account their sequential nature. During inference, our approach identifies which samples to take using k-space features, considering that one task (severity) will be following the confirmation of the disease.
• Figure 2: Illustration of the proposed sequential diagnostic active sampling. Initially, a randomly sampled $k$-space subset is fed into two pre-trained disease and severity classifiers. The active sampler takes as input the features extracted by the classifiers and decides the next $k$-space sampling location. After exhausting the sampling budget or meeting a user-defined criterion, the disease classifier confirms the presence of disease and subsequently the severity classifier is queried to estimate the severity level. During training, the active sampler is optimized with a novel step-wise weighting mechanism that combines the disease and severity rewards and enables a smooth transition.
• Figure 3: Step-wise weighting evolution for multi-objective diagnosis following lexicographic ordering. The step-wise weight dynamically determines the preference of the reward feedback in disease detection and severity quantification objectives, which ensures a smooth transition.
• Figure 4: Performance comparison of our proposed Weighted Policy and two single-task oriented policies. The horizontal axis indicates the cumulative lines acquired when sampling. The $80$-step active sampling procedure starts from $16$ randomly sampled $k$-space lines.
• Figure 5: Average $k$-space sampling behaviors across all subjects of our proposed Weighted Policy and benchmarks. A High value in the plot represents a high probability to be sampled. The horizontal axis indicates the cumulative lines acquired when sampling.