Deep Learning Assisted Outer Volume Removal for Highly-Accelerated Real-Time Dynamic MRI

TL;DR

This work tackles aliasing in highly accelerated real-time cine MRI by introducing outer volume removal (OVR) as a post-processing step. A DL-based ghosting detector estimates motion-induced, pseudo-periodic outer-volume artifacts from composite time-interleaved frames, enabling subtraction of the outer-volume signal from k-space. The outer-volume-subtracted data are then reconstructed with a physics-driven DL (PD-DL) network trained using a specialized OVR-aware loss and self-supervised SSDU, yielding high spatio-temporal fidelity at $R=8$ that rivals baseline $R=4$ acquisitions. The approach avoids acquisition changes, shows strong qualitative and quantitative improvements across retrospective and prospective datasets, and generalizes to different sequences and field strengths, with potential applicability to broader dynamic MRI challenges.

Abstract

Real-time (RT) dynamic MRI plays a vital role in capturing rapid physiological processes, offering unique insights into organ motion and function. Among these applications, RT cine MRI is particularly important for functional assessment of the heart with high temporal resolution. RT imaging enables free-breathing, ungated imaging of cardiac motion, making it a crucial alternative for patients who cannot tolerate conventional breath-hold, ECG-gated acquisitions. However, achieving high acceleration rates in RT cine MRI is challenging due to aliasing artifacts from extra-cardiac tissues, particularly at high undersampling factors. In this study, we propose a novel outer volume removal (OVR) method to address this challenge by eliminating aliasing contributions from non-cardiac regions in a post-processing framework. Our approach estimates the outer volume signal for each timeframe using composite temporal images from time-interleaved undersampling patterns, which inherently contain pseudo-periodic ghosting artifacts. A deep learning (DL) model is trained to identify and remove these artifacts, producing a clean outer volume estimate that is subsequently subtracted from the corresponding k-space data. The final reconstruction is performed with a physics-driven DL (PD-DL) method trained using an OVR-specific loss function to restore high spatio-temporal resolution images. Experimental results show that the proposed method at high accelerations achieves image quality that is visually comparable to clinical baseline images, while outperforming conventional reconstruction techniques, both qualitatively and quantitatively. The proposed approach provides a practical and effective solution for artifact reduction in RT cine MRI without requiring acquisition modifications, offering a pathway to higher acceleration rates while preserving diagnostic quality.

Figures (8)

• Figure 1: Illustration of the decomposition of a composite real-time cine cardiac MR image into various components. For simpler visualization, the heart is depicted as the only moving object, while surrounding tissues are treated as stationary, and $R = 4$ is used. Both the moving components and stationary components of each time-frame contribute to the composite image as foldovers with distinct modulation coefficients due to the shifted phase encoding lines across time-frames. The moving components add constructively at the true heart location, creating a temporally averaged representation, while their summation at the other foldover locations leads to a pseudo-periodic ghosting artifact due to cardiac motion between time-frames and the distinct modulation coefficients. Conversely, the stationary components add constructively at the central foldover while canceling out at the other foldover locations.
• Figure 2: Importance of accounting for ghosting artifacts during OVR. (a) TGRAPPA reconstruction at acceleration rate $R$ = 4. (b) CG-SENSE reconstruction of OVR k-space data at $R$ = 4, using composite coil images as the outer volume, without ghosting artifact correction. (c) CG-SENSE reconstruction of OVR k-space data at $R$ = 4, where the outer volume is estimated after ghosting artifact removal from composite images. The left column shows the baseline image during systole. The reconstruction of the same data in the middle column shows temporal blurring, effectively yielding a different cardiac phase due to the effect of the ghosting artifacts present in the data, while also exhibiting blurring artifacts. The right column matches the correct cardiac phase and shows high image fidelity.
• Figure 3: Proposed reconstruction pipeline: (a) DL-based ghosting detection: A ResNet with 15 residual blocks (RBs) takes four adjacent composite images, concatenating their real and imaginary components into eight input channels ($\textrm{C}_\textrm{in}=8$), and estimates the ghosting artifacts for the corresponding time frame $t_0$, producing eight output channels ($\textrm{C}_\textrm{out}=8$). The stationary background coil images, $\mathbf{x}_\textrm{background}$, are then obtained by subtracting the detected ghosting artifacts from the composite images of the target time frame. (b) Physics-driven DL-based unrolled network: The network consists of 35 unrolls, where each iteration block contains one ResNet and one data fidelity (DF) block. The ResNets ($\textrm{C}_\textrm{in}=2$, $\textrm{C}_\textrm{out}=2$, 15 residual blocks) perform the proximal operation for the regularizer on the previous DF block output, and share the same network weights ($\theta$). The DF block takes the zerofilled image ($\mathbf{x}_\textrm{0}$), the last ResNet output ($\mathbf{z}_\textrm{k}$), sensitivity maps, and the acceleration mask, producing a data-consistent image as output using conjugate gradient. (c) The ResNet structure is used in both the ghosting detection network and the proximal operators of the unrolled network. (d) Outer volume subtraction and PD-DL reconstruction: The background image is masked using the OVR mask ($\mathbf{m}_\textrm{OVR}$) and transformed into k-space via the Fourier transform ($\mathcal{F}^{\Omega_t}$), then subtracted from the acquired signal ($\mathbf{y}^{\Omega_t}$). The OVR k-space signal $\mathbf{y}_\textrm{OVR}^{\Omega_t}$ is mapped back to the image domain using the adjoint encoding operator $( \mathbf{E}^{\Omega_t} ) ^H$ and fed into the unrolled network. As expected, the network output exhibits minimal background signal. (e) Final image formation by combining the masked background with the reconstruction.
• Figure 4: PD-DL reconstruction of the OVR k-space using (a) masked sensitivity maps, (b) full sensitivity maps, and (c) full sensitivity maps with consistency through the proposed loss function in \ref{['full_cost']}. The left column shows artifacts, while the middle column exhibits signal loss. The proposed consistency mechanism effectively mitigates both issues, as demonstrated in the right column.
• Figure 5: Detection of the ghosting artifact of a composite RT cine image (retrospective acceleration $R=8$). The composite image (left) contains both static and dynamic structures, with motion-induced artifacts marked by red arrows, and the reference ghosting map (middle) is generated using TGRAPPA at $R=4$. The proposed method (right) effectively isolates ghosting artifacts, capturing their pseudo-periodic patterns. We also note that our method benefits from higher-SNR composite images, compared to the noisier reference labels from TGRAPPA ($R=4$), which suffer from spatially varying g-factor noise. As a result, the DL-based estimation represents a denoised version of the reference ghosting map, as expected.