Flow-Guided Implicit Neural Representation for Motion-Aware Dynamic MRI Reconstruction

TL;DR

This paper tackles dynamic MRI reconstruction under severe undersampling by jointly modeling the image sequence and its motion using two coupled implicit neural representations. The image INR $I(\theta)$ and the flow INR $\mathcal{G}(\phi)$ are linked via the optical-flow constraint, with data fidelity and TV-based regularization enforcing physical plausibility and smoothness. The method operates in a fully unsupervised manner, requiring no external training data or pre-estimated flow, and demonstrates superior reconstruction quality, motion estimation accuracy, and temporal fidelity on cardiac cine datasets compared to state-of-the-art baselines. The approach holds potential for high-acceleration dynamic imaging by enabling continuous, physics-informed reconstruction of both images and motion fields.

Abstract

Dynamic magnetic resonance imaging (dMRI) captures temporally-resolved anatomy but is often challenged by limited sampling and motion-induced artifacts. Conventional motion-compensated reconstructions typically rely on pre-estimated optical flow, which is inaccurate under undersampling and degrades reconstruction quality. In this work, we propose a novel implicit neural representation (INR) framework that jointly models both the dynamic image sequence and its underlying motion field. Specifically, one INR is employed to parameterize the spatiotemporal image content, while another INR represents the optical flow. The two are coupled via the optical flow equation, which serves as a physics-inspired regularization, in addition to a data consistency loss that enforces agreement with k-space measurements. This joint optimization enables simultaneous recovery of temporally coherent images and motion fields without requiring prior flow estimation. Experiments on dynamic cardiac MRI datasets demonstrate that the proposed method outperforms state-of-the-art motion-compensated and deep learning approaches, achieving superior reconstruction quality, accurate motion estimation, and improved temporal fidelity. These results highlight the potential of implicit joint modeling with flow-regularized constraints for advancing dMRI reconstruction.

Figures (7)

• Figure 1: Framework of the proposed model. Spatiotemporal coordinates are input into two independent INR networks with identical architectures (INR Net A is shown as an example in the figure, which is processed through Hash encoding to serve as input for the MLP). The INR Net A outputs dual-channel image intensity, which is combined with real and imaginary parts to generate the reconstructed image. The reconstructed image undergoes undersampled FFT and sensitivity map modulation to obtain predicted k-space data, which is compared with real k-space data to calculate data consistency loss. The INR Net B outputs predicted optical flow vectors, which are orthogonalized with image gradients to obtain the optical flow regularization loss. Total variation is applied to both the output image sequence and optical flow vectors to constrain spatial noise. The loss function is minimized to iteratively update the parameters of the two INR networks.
• Figure 2: Frameworks of ablation study. (a) compute the gradient directly on the reconstructed MRI without utilizing the continuity of INR, and (b) pre-estimate the optical flow using low-frequency signals in the K-space (fully sampled 32×32 center region) without joint reconstruction.
• Figure 3: Reconstruction results of the proposed method and baseline methods. Visualization of reconstruction results at Random Cartesian (AF=8), and VISTA (AF=8, 24) undersampling masks. The second row shows the magnified region of the heart (orange box). The third row displays the x-t image outlined by the blue line (the 120nd slice). The fourth and fifth rows respectively show the error maps. The red arrow highlights the advantages of our method.
• Figure 4: Temporal interpolation performance comparisons. The orange color indicates the frames generated by INR interpolation.
• Figure 5: Reconstruction results and optical flow estimation under different AF on simulated samples (Cheat-GT). The left side represents systole, while the right side represents diastole. The first row shows Cheat-GT and the reconstruction results of different AFs. The second and third rows show the motion field and error map respectively. The top-right corner shows a color wheel representing the motion field.