Adaptive Plane Reformatting for 4D Flow MRI using Deep Reinforcement Learning

TL;DR

<3-5 sentence high-level summary> AdaPR addresses the challenge of time-consuming and observer-dependent plane reformatting in 4D flow MRI by introducing an adaptive plane reformatting framework based on deep reinforcement learning with a local coordinate system. Implemented with A3C, AdaPR navigates volumes of arbitrary position and orientation, achieving robust plane localization and orientation with mean angular error ~6.3° and distance error ~3.4 mm, and flow measurements highly concordant with manual observers (R^2 ≈ 0.97–0.98). Across 88 multi-vendor datasets, AdaPR demonstrated strong generalization to different scanners and anatomies, with performance remaining stable under orientation and position perturbations. These results suggest AdaPR as a practical, scalable solution for automated plane reformatting in 4D flow MRI and potentially other imaging modalities requiring flexible view planning.

Abstract

Background and Objective: Plane reformatting for four-dimensional phase contrast MRI (4D flow MRI) is time-consuming and prone to inter-observer variability, which limits fast cardiovascular flow assessment. Deep reinforcement learning (DRL) trains agents to iteratively adjust plane position and orientation, enabling accurate plane reformatting without the need for detailed landmarks, making it suitable for images with limited contrast and resolution such as 4D flow MRI. However, current DRL methods assume that test volumes share the same spatial alignment as the training data, limiting generalization across scanners and institutions. To address this limitation, we introduce AdaPR (Adaptive Plane Reformatting), a DRL framework that uses a local coordinate system to navigate volumes with arbitrary positions and orientations. Methods: We implemented AdaPR using the Asynchronous Advantage Actor-Critic (A3C) algorithm and validated it on 88 4D flow MRI datasets acquired from multiple vendors, including patients with congenital heart disease. Results: AdaPR achieved a mean angular error of 6.32 +/- 4.15 degrees and a distance error of 3.40 +/- 2.75 mm, outperforming global-coordinate DRL methods and alternative non-DRL methods. AdaPR maintained consistent accuracy under different volume orientations and positions. Flow measurements from AdaPR planes showed no significant differences compared to two manual observers, with excellent correlation (R^2 = 0.972 and R^2 = 0.968), comparable to inter-observer agreement (R^2 = 0.969). Conclusion: AdaPR provides robust, orientation-independent plane reformatting for 4D flow MRI, achieving flow quantification comparable to expert observers. Its adaptability across datasets and scanners makes it a promising candidate for medical imaging applications beyond 4D flow MRI.

Figures (6)

• Figure 1: Overview of the adaptive plane-reformatting framework (AdaPR) and reformatted planes.A. Asynchronous Advantage Actor–Critic (A3C) architecture: An initial stack of volumes goes through three 3D convolutional layers and 1 fully-connected layer. The output of this layer feeds an LSTM cell with the previous LSTM hidden and cell state. Next, the Actor layers compute $\mu_a$ and $\sigma_a$ to estimate the policy $\pi(s,a)$ and the Critic layer estimates the value function $G(s)$. The model parameters are updated with equation (6) and the policy is sampled to transition to a new state. B. Example of manual reformatted planes used for training. C. Diagram of ResNet blocks in convolutional layers.
• Figure 2: Sensitivity of plane-reformatting accuracy to rigid transformations of the input volumes. Heatmaps summarize average performance across all test volumes and planes when the same scans are perturbed by rotations of 5°, 15°, 25° and translations of 5, 15, 25 mm along all axes. Top panels report angular error (°); bottom panels report distance error (mm), both computed between the automated planes and the reference manual planes of observer 1. AdaPR maintains a near-constant error relative to the average error without rotations or translations, whereas VanillaPR variants show progressive degradation with larger perturbations.
• Figure 3: Distribution of plane reformatting metrics for AdaPR. Whiskers range from the 2.5th to the 97.5th percentile. The line represents the median, and the cross represents the mean. (A) Angular error ($^\circ$). (B) Distance error (mm). **, ***, and **** represent ART-C post-hoc contrasts with significance levels $<10^{-2}$,$<10^{-3}$, and $<10^{-4}$, respectively.
• Figure 4: Examples of plane reformatting. Planes placed by observer 1 (O1), observer 2 (O2), and AdaPR planes are shown in yellow, green, and red, respectively. The left-hand column depicts a healthy volunteer; the right-hand column shows a patient with a bicuspid aortic valve (BAV).
• Figure 5: Flow measurements (L/min) at four vessels (AAo: ascending aorta; PA: pulmonary artery; RPA: right pulmonary artery; LPA: left pulmonary artery) for AdaPR, Observer 1, and Observer 2. Box plots show median (line), mean (+), interquartile range (box),and whiskers range from the 2.5th to the 97.5th percentile. Different letters represent groups; box plots with the same letter have no significant differences based on two-way ANOVA post-hoc comparisons (p $>$ 0.05).