Abdominal MR Multitasking for radiotherapy treatment planning: A motion-resolved and multi-contrast 3D imaging approach

Abstract

Purpose:

Radiotherapy treatment planning (RTP) using MR has increasingly been used for the abdominal site. Multiple contrast weightings and motion-resolved imaging are desired for accurate delineation of the target and various organs-at-risk (OARs) and patient-tailored planning. Current MR protocols achieve these through multiple scans with distinct contrast and variable respiratory motion management strategies and acquisition parameters, leading to a complex and inaccurate planning process. This study presents a standalone MR Multitasking (MT)-based technique to produce volumetric, motion-resolved, multi-contrast images for abdominal RTP.

Methods:

The MT technique resolves motion and provides a wide range of contrast weightings by repeating a magnetization-prepared (saturation-recovery and T2 preparations) spoiled gradient echo readout series and adopting the MT image reconstruction framework. The performance of the technique was assessed through digital phantom simulations and in-vivo studies of both healthy volunteers and patients with liver tumors.

Results:

In the digital phantom study, the MT technique presented structural details and motion in excellent agreement with the digital ground truth. The in vivo studies showed that the motion range was highly correlated (R² = 0.82) between MT and 2D cine imaging. MT allowed for a flexible contrast-weighting selection for better visualization. Initial clinical testing with inter-observer analysis demonstrated acceptable target delineation quality (Dice coefficient = 0.85 ± 0.05, Hausdorff Distance = 3.3 ± 0.72 mm).

Conclusion:

The developed MR Multitasking-based, abdomen-dedicated technique is capable of providing motion-resolved, multi-contrast volumetric images in a single scan, which may facilitate abdominal radiotherapy treatment planning.

1. Introduction

Radiation therapy, or radiotherapy, is a highly effective cancer treatment and has been used in the management of more than half of patients with cancer(1). External beam radiotherapy, the most common type of radiotherapy, aims the prescribed radiation dose at the tumor while sparing normal tissue as much as possible. Its success heavily relies on careful radiotherapy treatment planning (RTP) through an imaging session in which the patient is positioned in a way simulating the treatment setup and is scanned for a better understanding of individual anatomies(2). CT is routinely used for this purpose because of its wide accessibility, high spatial resolution, and capability to provide the electron density information needed for dose calculation(3). However, CT often offers inadequate soft-tissue contrast, particularly at the brain, head-neck, and abdomen sites. For better delineation of the treatment target and organs-at-risk (OARs), MR has increasingly been adopted as an imaging platform complementary to CT(4–7). More recently, with the introduction of MR-guided radiation therapy systems (6,8–12), MR-only based RTP has been under intensive investigation to mitigate the drawbacks (such as misregistration-induced geometric errors, additional cost, and logistics) associated with the hybrid CT-MR based RTP workflow(5).

Conventional MR imaging session for abdominal RTP occurs days before treatment session and the MR RTP imaging protocols consist of multiple sequential scans to produce different contrast weightings, such as T1-weighted (T1w) and T2-weighted (T2w), to facilitate accurate delineation of the treatment target and various OARs (13,14). In this context, routinely used MR pulse sequences, which are typically adapted from diagnostic protocols, often use different respiratory motion management strategies - respiratory gating or triggering, free-breathing data averaging, or breath holding. Consequently, acquired images are often inter-scan mismatched in the respiratory motion phase, or limited to certain phases (end-expiration or end-inspiration). Moreover, the variations in acquisition parameters, such as spatial resolution and coverage, and inter-scan bulk movement may contribute to the complexity and inaccuracy of MR RTP.

Over the past decade, respiratory motion phase-resolved MR techniques with volumetric spatial coverage, also known as 4D MR, have been investigated to better serve the abdominal RTP application by providing complete respiratory motion characterization(15). Hu et al. (16), Cai et al.(17), and Du et al. (18) developed slice-sorting algorithms to generate 4D MR images based on an internal or external respiratory motion surrogate. However, stacking 2D slices to produce a 3D volume of a specific respiratory motion phase is subject to stitching artifacts when patients experience irregular breathing patterns during image acquisition. Subsequently, several groups developed 3D acquisition-based 4D MR techniques. Deng et al.(19), Rank et al.(20), and Feng et al.(21) used navigator signals periodically extracted from the acquired k-space data to sort the imaging data into respiratory bins. Huttinga et al. (22,23) developed a motion model-based 4D MRI approach whereby a reference 3D image and estimated motion vector fields are used to generate images corresponding to different motion phases. However, all the techniques mentioned above are limited to a single contrast weighting, typically T1 weighting, due to the use of steady-state free precession acquisitions for achieving fast imaging. Additional scans for direct acquisition or model-based generation(24) of other contrast weightings subject the final set of images to inter-scan misalignment.

This work aimed to develop an abdomen-dedicated MR technique for RTP by providing motion-resolved volumetric images with a wide range of contrast weightings through a single scan. The technique builds upon MR Multitasking (MT)(25), a low-rank tensor framework, that exploits correlations in images across motion phases and along the dynamic signal evolution and thus achieves motion resolving and contrast-weighting flexibility at the same time. In addition to temporally averaged motion phase-resolved images, the technique can also produce retrospective real-time images, a useful capability for interrogating breathing-pattern changes(26). The proof-of-concept of the MT technique was demonstrated in a digital phantom and human volunteers.

2. Methods

2.1. Pulse Sequence Design

The MT sequence achieves flexible contrast weightings by repeating a dynamic signal evolution as divided into three stages (Figure 1A): “T1w Segment” where a saturation-recovery (SR) radiofrequency (RF) pulse is followed by a series of spoiled gradient echo (GRE) readouts to generate T1w contrast, then “Low flip-angle (FA) Segment” where a series of low-FA spoiled GRE readouts are subsequently applied to allow the longitudinal magnetization to recover while continuing signal acquisition at proton density-weighted (PDw) contrast, and finally “T2w Segment” where T2-prepared spoiled GRE readouts proceed to the end to generate T2w contrast.

FIGURE 1.

(A) The pulse sequence diagram of the proposed MT-RTP technique. The sequence builds on continuous spoiled gradient echo readouts with a repetitive three-stage signal evolution, including the saturation-recovery (SR) prepared T1-weighted (T1w) segment, the low flip-angle (FA) segment for proton density-weighted (PDw) contrast, and the T2 prepared (T2prep) T2-weighted (T2w) segment. Periodically, one training data readout (red arrows) is interleaved with other 9 imaging data readouts. (B) Illustration of the Cartesian spiral-in trajectory with Cartesian readout in k_x and spiral interleaving in the k_y-k_z plane. Each spiral interleaf has 10 readouts, with the last readout located at k_y = k_z= 0 serving as the subspace training data.

The k-space sampling trajectory (Figure 1B) follows a Cartesian spiral-in pattern in the $k_{\mathrm{y}}-k_{\mathrm{z}}$ plane(27,28). Each spiral interleaf, consisting of 10 readouts, traverses from the periphery of the $k_{\mathrm{y}}-k_{\mathrm{z}}$ plane to the center $\left(k_{\mathrm{y}}=k_{\mathrm{z}}=0\right)$ with increasing sampling densities. The last readout at the center serves as the subspace training data as elaborated later. Consecutive spiral interleaves shift by a golden angle. The above sampling strategy is intended to ensure: 1) computationally efficient image reconstruction with a 3D Cartesian acquisition, 2) A pseudo-random k-space sampling for uniform image quality along each time dimension, and 3) small k-space jumps and resultant reduced eddy current effects, particularly for subspace training data readouts.

The sequence adopts water-excitation RF pulses for GRE acquisitions with TR/TE = 6.0/2.3 ms, FA of T1w/Low FA/T2w Segment = 3°/1°/5°, bandwidth = 762 Hz/pixel. A BIREF-4(29) T2 preparation with a duration of 56 ms is used. A complete dynamic signal evolution cycle lasts 3.2 s.

2.2. Image Reconstruction

Multi-Dimensional Image Factorization:

The multi-dimensional MR image can be represented as a tensor:

$$𝒜\in {\mathbb{C}}^{N_x{N}_y{N}_z\times N_c\times N_m},$$

where $N_x,N_y,N_z$ are the three spatial dimensions, and $N_c$ and $N_m$ are the number of contrasts and number of respiratory motion states, respectively. This image tensor can be subsequently factorized by matrix multiplications:

$$𝒜=\Phi {\times }_1{\mathit{U}}_{\mathit{r}}$$ (1)

where ${\mathbf{U}}_{\mathbf{r}}\in {\mathbb{C}}^{{L\times N}_x{N}_y{N}_z}$ is the spatial basis and $L$ is the low-rank dimension of the spatial basis. The tensor $\mathrm{\Phi }\in {\mathbb{C}}^{L\times N_c\times N_m}$ is itself a decomposable low-rank tensor containing temporal basis functions and core tensor weightings(25).

Calculation of $\mathrm{\Phi }$

The subspace training data acquired at the k-space center is denoted as ${\mathit{d}}_{tr}\in {\mathbb{C}}^{N_x{N}_{coil}\times (N_t/10)}$ where $N_{coil}$ is the number of coils and $N_t$ is the total number of image lines acquired ($N_t/10$ is the number of navigator lines that are acquired every 10 readout lines). After motion binning (described later), we can perform tensor nuclear norm optimization to extract a training data tensor ${𝒟}_{tr}\in {\mathbb{C}}^{N_x{N}_{coil}\times \left(N_t/10\right)\times N_m}$:

$${\widehat{𝒟}}_{tr}=\mathit{arg}\underset{{𝒟}_{\mathit{tr}}}{\mathit{min}}{‖{\mathit{d}}_{tr}-{\Omega }_{tr}\left({𝒟}_{tr}\right)‖}_2^2+\lambda {\sum }_i{‖{\mathit{D}}_{tr\left(i\right)}‖}_{\mathbf{*}}$$ (2)

where ${\Omega }_{tr}$ is the sampling operation on the tensor and where the subscript ($i$) means mode-$i$ unfolding. A short length $\widehat{\Phi }\in {\mathbb{C}}^{L\times \left(N_t/10\right)\times N_m}$ can then be extracted from ${\widehat{𝒟}}_{tr}$ using higher order singular value decomposition(30). A simple interpolation using piecewise cubic polynomial along the signal evolution dimension will give rise to the full length $\Phi \in {\mathbb{C}}^{L\times N_c\times N_m}$.

Calculation of ${\mathbf{U}}_{\mathbf{r}}$

With the calculated $\Phi ,{\mathit{U}}_{\mathit{r}}$ can be derived by solving the following optimization problem:

$${\widehat{\mathit{U}}}_r=\mathit{arg}\underset{{\mathit{U}}_{\mathit{r}}}{\mathit{min}}{‖\mathit{d}-\Omega \left(\mathrm{\Phi }{\times }_1\mathit{F}\mathit{S}{\mathit{U}}_{\mathit{r}}\right)‖}_2^2+R\left({\mathit{U}}_{\mathit{r}}\right)$$ (3)

where $\mathit{d}\in {\mathbb{C}}^{N_x{N}_{coil}\times Nt}$ is the full set of acquired data, $\Omega$ is the sampling trajectory operator, $\mathit{S}$ is the coil sensitivity operator, and $\mathit{F}$ is Fourier transform. The term $R\left({\mathit{U}}_{\mathit{r}}\right)$ denotes a sparsifying regularization, which, in this case, is wavelet regularization.

Binning

Because we acquire signals with changing contrast, it is challenging to perform respiratory binning directly based on the subspace training data. Therefore, we developed an image-based method where crude real-time images with changing contrast are first reconstructed, and the liver dome position is then used for respiratory motion binning. The process is as follows:

Approximating ${\mathbf{d}}_{\mathrm{t}\mathrm{r}}$ with rank L after SVD, we have:

$$d_{\mathrm{t}\mathrm{r}}\approx {\mathrm{U}}_L{\mathrm{\Sigma }}_L{V}_L^H$$ (4)

A real-time temporal basis ${\widehat{\mathit{\Phi }}}_{rt}\in {\mathbb{C}}^{L\times (N_t/10)}$ can be calculated by:

$${\widehat{\mathit{\Phi }}}_{rt}=V_L^H$$ (5)

After a simple interpolation, we can recover the full-length temporal basis function ${\Phi }_{rt}\in {\mathbb{C}}^{L\times Nt}$. With an optimization step without any regularization, an initial spatial factor ${\mathit{U}}_{\mathbf{0}}{\in \mathbb{C}}^{N_x{N}_y{N}_z\times L}$ can be calculated by:

$${\widehat{\mathit{U}}}_0=\mathit{arg}\underset{{\mathit{U}}_{\mathbf{0}}}{\mathit{min}}{‖\mathit{d}-\Omega \left(\mathit{F}\mathit{S}{\mathit{U}}_{\mathbf{0}}{\mathit{\Phi }}_{\mathit{r}\mathit{t}}\right)‖}_2^2$$

A crude real-time image series with changing contrast corresponding to the navigator lines is then produced by:

$${𝒜}_0={\mathit{U}}_{\mathbf{0}}{\widehat{\mathit{\Phi }}}_{\mathit{r}\mathit{t}}$$ (6)

With morphological image open operation (31) to smooth the organ structure and basic thresholding, the real-time liver dome position can be determined. This process is also demonstrated by supporting information video S1.

Reconstruction Environment

Reconstruction algorithms were implemented in MATLAB (2020 a, MathWorks, MA). Reconstruction time was approximately 10 minutes when using a Linux server with 96 core intel Xeon CPUs, Nvidia GeForce 3090 (24GB) and 256GB RAM.

2.3. Digital Phantom Experiments

The XCAT digital phantom model (32) was used to generate simulated k-space data in accordance with the tissue T1/T2 values at 3T and the proposed signal evolution. The simulation platform was modified from an abdomen signal simulation package(33). Three types of respiratory cycles of different durations, ranges of superior-inferior (SI) diaphragm displacement, and ranges of anterior-posterior (AP) expansion were created: regular (4.2 s, AP 5 mm, SI 20 mm), long (6.3 s, AP 5.75 mm, SI 23 mm), and short (3.2 s, AP 4.5 mm, SI 18 mm). These types of cycles were mixed to simulate a complex breathing pattern according to an occurrence rate of 80%, 13.3%, and 6.7%, respectively. Major acquisition parameters included: coronal orientation, field-of-view = 256 (SI) × 384 (left-right) × 256 (AP) mm³, matrix size = 160 × 224 × 80, resolution = 1.6 × 1.6 × 3.2 mm³ (isotropic 1.6 mm after interpolation), scan time = 8 min. The rank for reconstruction is set to be L = 12.

To validate the motion-resolving capability, the reconstructed 4D MR images were compared with the digital ground truth at end-expiration and end-inspiration phases, respectively, for T1w, PDw, and T2w contrasts through the structural similarity (SSIM) index. Retrospective real-time T1w and T2w contrast-frozen images, generated based on the method proposed by Han et al.(26), were also compared with the real-time ground truth, for an arbitrarily selected cycle of each breathing type, through the SSIM index and the cross-correlation in the liver dome position.

2.4. In Vivo Experiments

All experiments were conducted with the approval of local Institutional Review Boards. Twenty-two healthy volunteers and 5 patients with liver tumors were recruited with informed consent and scanned using 3T MR scanners (MAGNETOM Vida, Siemens Healthineers, Erlangen, Germany) with an 18-channel surface coil and an integrated spine coil. All volunteers were instructed to lie supine with arms raised above the head during scanning to accord with the RTP practice. The proposed MT sequence and the rank L used for reconstruction were the same as the protocol described in the digital phantom experiment section above. Both MT-based multi-contrast 4D MR images and retrospective real-time images were generated.

In all healthy volunteers, real-time 2D images using a spoiled GRE sequence (coronal orientation, TR/TE = 3.6/1.3 ms, FA = 9°, resolution = 1.6 × 1.6 mm², slice thickness = 6 mm, temporal resolution = 300 ms, scan time = 1 min) were acquired as references. In 7 of them, a commercial T1w 4D MR sequence, i.e., StarVIBE (free breathing VIBE with stack-of-stars acquisition)(34), was acquired using the following parameters: coronal orientation, resolution = 1.6 × 1.6 × 3.0 (interpolated from 6.0) mm³, water excitation RF, TR/TE = 6.0/1.35 ms, FA = 9°, 8 respiratory phases, scan time = 9 min. A total of eight phases were sorted based on magnitude percentile.

The patients were scanned on an RTP-dedicated 3T MR system with a hard tabletop and a coil bridge. In addition to the MT sequence, the following reference sequences were acquired: breath-hold multi-slice T2 BLADE sequence (1.4 × 1.4 mm² transverse view in-plane resolution, 5 mm slice thickness, 1 mm slice gap, scan time = 3 mins); StarVIBE (1.3 × 1.3 mm² transverse view in-plane resolution, 3.75 mm slice thickness after interpolation, 8 respiratory phases, scan time = 9 mins).

For all 22 healthy volunteers, the respiratory motion range was measured at the liver dome for the real-time 2D reference and the MT-based 4D MR images (Supporting Information Video S2), respectively, and linear correlation was calculated. For the 7 volunteers who underwent StarVIBE scans, the signal-to-noise ratio (SNR) of liver parenchyma was calculated for both StarVIBE and MT-based T1w 4D MR images. Additionally, three pairs of T1w coronal slices were selected from the two scans for both end-expiration and end-inspiration phases. These images were blindly evaluated for visualization quality in a randomized order by an experienced reader using a 4-point scale (0: poor, 1: fair, 2: good, and 3: excellent). Wilcoxon signed-rank tests were performed to determine the difference in image quality.

For the 5 patients, we evaluated the ability of the proposed MT-based method to help accurately delineate the tumor targets and OARs by analyzing the inter-observer variability on the contours of the tumor target and select OARs. One radiologist and one radiation oncologist were tasked to independently prescribe contours for the tumor in the liver and two OARs (the right kidney and the spleen) at both end-inspiration and end-expiration phases by simultaneously referring to T1w, PDw and T2w images. The inter-observer agreement in contours was analyzed using the Dice coefficient and 95 percentile Hausdorff distance (HD95). Additionally, motion ranges of the tumor contours (center of mass) from the proposed MT-based method were benchmarked against those from StarVIBE in three directions (SI, AP, LR).

All statistical analyses were performed using MATLAB. A p value less than 0.05 indicated a significant difference.

3. Results

3.1. Digital Phantom Experiments

Figure 2 shows the reconstructed MT-based T1w, PDw, and T2w 4D MR images at the end-inspiration and end-expiration phases and their corresponding digital ground truths. Figure 3 shows the progression of the SSIM index across contrasts for end-inspiration and end-expiration phases, respectively. Both qualitative and quantitative comparisons indicated that the proposed MT technique largely preserved the anatomical structure and their contrasts, although T2w images were associated with slightly reduced image quality and structural integrity. SSIM indices for contrasts immediately after the SR pulse are high, largely due to inherently low signal intensity for both the reconstructed image and the ground truth. Figure 4 demonstrates that our retrospectively reconstructed real-time images were accurate in depicting motion and anatomy irrespective of time-varying breathing cycles.

FIGURE 2.

T1-, proton-density (PD)-, and T2-weighted digital phantom images acquired using the MT sequence and their corresponding digital ground truths for both end-expiration (EE) and end-inspiration (EI) phases.

FIGURE 3.

Progression of SSIM index compared to the ground truth across contrasts for both end-inspiration and end-expiration from digital phantom simulations.

FIGURE 4.

Quantitative analysis of retrospectively reconstructed real-time T1-weighted (left) and T2-weighted (right) contrasts for three breathing pattern types: normal (4.2s), long (6.3s), and short (3.2s). Top row shows the cross-correlation in the liver dome position between the contrast-frozen MT and the digital ground truth. Bottom row shows the SSIM index between them.

3.2. In Vivo Experiments

The MT technique was able to generate motion-resolved, multi-contrast volumetric images (Figure 5). The number of motion phases for all subjects ranged from 7 to 11 depending on different respiratory motion ranges. Users were essentially allowed to freely view anatomies at any available contrast and slice in a phase-resolved (i.e., 4D MR) or real-time fashion. More detailed video demonstrations of multi-contrast 4D MR are included in supporting information video S3. A video of retrospective multi-contrast real-time images is demonstrated in supporting information video S4.

FIGURE 5.

Demonstration of the MT images across contrasts and motion phases from a healthy volunteer.

Figure 6 shows the motion range correlation between MT images (T1w) and 2D real-time reference image across 22 volunteers. The fitted line was calculated to be y = 0.95x +2.6, with R² = 0.82.

FIGURE 6.

Respiratory motion ranges of 22 healthy volunteers measured at the liver dome from MT images and the 2D real-time reference.

Figure 7A shows a side-by-side comparison between StarVIBE and MT images. Both sequences resolved respiratory motion. However, MT allowed for selection of a T1 contrast weighting at the user’s discretion, potentially affording a visually better depiction (such as SNR measured from T1w images at 600 ms following the SR RF pulse, shown in Figure 7B) of anatomies than StarVIBE. MT was also capable of producing additional PDw and T2w contrasts. For the end-expiration phase, the image quality scores of MT T1w images were significantly higher than that of StarVIBE (2.52 ± 0.60 vs. 2.09 ± 0.30, p = 0.011). For the end-inspiration phase, image quality scores were comparable between the two techniques (1.95 ± 0.38 vs. 1.80 ± 0.68, p = 0.436).

FIGURE 7.

(A) Representative healthy volunteer images obtained by the clinically available StarVIBE sequence and the proposed MT technique. MT produces motion-resolved images with flexible, multiple imaging contrasts. (B) SNR comparison between the T1-weighted MT images and StarVIBE images of 7 volunteers.

Figure 8 shows the comparison of MT images and the reference StarVIBE and T2 BLADE images from a patient with cholangiocarcinoma. MT images provided comparable image contrast to the reference sequences. However, the inherently co-registered multi-contrast volumetric images from the MT technique provided superior 3D visualization of the tumor.

FIGURE 8.

Comparison of MT images (both T1-weighted and T2-weighted) with the reference free-breathing StarVIBE (T1-weighted) and breath-hold T2-weighted BLADE sequences in a patient with cholangiocarcinoma. Images are shown in both coronal and transverse views. Tumor locations are marked by arrows.

Figure 9 shows the quantitative analysis on the inter-observer agreement in the contours of the tumor, the right kidney and the spleen at both end-expiration and end-inspiration. The average Dice coefficients are 0.85 ± 0.05, 0.93 ± 0.02, and 0.95 ± 0.01 and the average HD95 are 3.3 ± 0.72, 2.1 ± 0.32 and 1.7 ± 0.21 mm, respectively.

FIGURE 9.

Dice coefficients and Hausdorff Distance (95 percentile) of segmentation results from two reviewers using MT images. Results for end-expiration (top) and end-inspiration (bottom) are shown.

Table 1 summarizes the motion ranges of the tumor contours (center of mass) from the proposed MT-based method and StarVIBE in three directions (SI, AP, LR) from the 5 patients.

TABLE 1.

Summary of the motion ranges of the tumor contours (center of mass) derived from the proposed MT-based method and from StarVIBE in three directions (SI, AP, LR) for the 5 patients.

	SI Motion Range (mm)		AP Motion Range (mm)		LR Motion Range (mm)
	MT-MR	StarVIBE	MT-MR	StarVIBE	MT-MR	StarVIBE
Patient 1	9.3	9.0	1.5	2.8	1.9	0.7
Patient 2	9.2	11.0	−0.1	2.1	0.1	1.9
Patient 3	7.7	7.8	2.9	2.9	0.8	1.5
Patient 4	10.6	0.1	2.3	−0.2	−0.3	0.1
Patient 5	0.9	0.4	0.7	0.9	−2.1	0.5

4. Discussion

In this work, we developed a motion-resolved, multi-contrast 3D MR technique based on MR Multitasking for abdominal RTP. In comparison with digital ground truths and reference 2D and 4D MR methods, the MT technique is capable of accurately depicting anatomical structures and respiratory motion. It is also advantageous over the existing clinical protocol, consisting of T1w 4D MR (e.g., StarVIBE) and other T2w/PDw sequences (e.g., 2D BLADE), by providing better SNR and image quality scores at T1w contrast, additional inherently co-registered contrast weightings, consistent imaging parameters among various contrast weightings, and a shorter scan time. Additionally, the MT technique uniquely offers retrospective real-time images, thus providing insights into respiratory pattern changes during the planning scan. With all the rich information, the MT technique may drastically simplify the abdominal MR-based RTP workflow.

The feasibility of this technique was initially demonstrated on a digital phantom with computer simulations. This offered the ground truth for anatomic structures and motion and allowed us to benchmark the performance of the proposed signal evolution profile and reconstruction. The digital simulation results showed that the proposed technique produced acceptable image quality for all contrasts and motion phases in the presence of random, varying respiratory durations and amplitudes. However, the T2w contrast exhibited slightly reduced image quality and structural integrity, which is likely attributed to the more transient nature of the T2w contrast compared to T1w and PDw contrast. These findings were also consistent with the results from in vivo experiments. In in vivo cases, unexpected movements (e.g., twitching, stretching) during the scan can further compromise the image quality for the T2w contrast due to its more transient signal evolution. In addition to the constraints arising from limited sampled data, the quality of T2w images also suffers in in vivo cases because T2-preperation modules are susceptible to field inhomogeneity, which is common at 3T and given the large field-of-view in abdominal imaging. In this study, we attempted to address the issue of inhomogeneity by employing BIREF-4(29) refocusing pulses and B₀ field shimming. It is important for future research to investigate alternative methods for generating T2w contrast that are more robust to field inhomogeneity and have longer-lasting effects. One potential approach to consider is the combination of GRE and TSE as proposed by Benkert et al(35). Additionally, we are currently investigating a new type of RF coil that incorporates B0 shimming capability, enabling inherent field homogenization, and eliminating the need for sequence modifications(36).

In this work, we adopt a straightforward way to sort respiratory motion under real-time changing contrast by performing a quick initial reconstruction and then applying morphological image processing. Essentially, our binning technique is based on tracking the position of the liver dome by thresholding the image following the morphological image open operation(31). Although the MT results shown in this paper is based on binning with absolute liver dome position, the real-time dome position can be further exploited to enable bi-directional binning, separating inspiration and expiration phases. This potential feature will therefore enable hysteresis analysis. Our proposed technique does not rely on the assumption that the respiratory motion is strictly smooth and periodic as is also demonstrated with the digital phantom simulation results. However, the time-varying contrast may result in uncertainty of the liver-lung boundary identified by thresholding, usually 1–2 pixels. Additionally, we observe that research subjects tend to stay in end-expiration phase longer than other transitory phases, given that they are at a relaxed state lying inside the scanner. As a result, we see that often the images at end-expiration have the sharpest images while other respiratory phases are blurrier. As an option in our reconstruction, Deformable vector fields (37) can be exploited to enhance image quality for some of respiratory phases using the end-expiration phase (See the example shown in Supporting Information Figure S1).

Our model assumes low rankness of the acquired signal. This requires both signal evolution and respiratory motion positions to be repetitive, if not strictly periodic. The MR-Multitasking framework does provide a mechanism for detecting sudden, unexpected but recoverable motions such as deep breaths and coughs, as previously demonstrated by Han et al.(26) and Wang et al.(38). If such motion is detected, the k-space data associated with the time period can be simply weighted down or removed. Given these sudden and unexpected motions are transient, their final impact on the images can be minimal. However, any bulk movement such as lateral movement during the scan will break the low-rank assumption and impact image quality. Nevertheless, patient movement restriction by using a stabilization device can be considered for those patients who have trouble staying still.

We utilized that fact that the multidimensional image can be expressed as a linear multiplication between a spatial basis and a temporal basis function, with the rank L = 12 chosen to be sufficiently high for separation between different dynamics such as T1w/T2w contrast evolution and respiratory motion. The real-time temporal basis function is extracted from the subspace training data. This alone allows the real-time volumetric display of the abdomen with changing contrasts. What this study provides in addition is the expression in the temporal basis subspace of the entire signal evolution curve for each respiratory motion state. As a result, this information can be used in conjunction with the real-time display mechanism to enable simultaneous multi-contrast real-time display of the abdomen, as already demonstrated by another study (26). Therefore, the MT technique can be employed in an MR-guided adaptive radiotherapy platform on the treatment day; a pre-beam scan will generate motion-resolved, multi-contrast volumetric images for radiation plan adaptation to the anatomy-of-the-day, and a beam-on scan will produce simultaneous multi-contrast real-time display for radiation beam guidance.

With an example from a patient with cholangiocarcinoma, we demonstrated that the MT technique was able to better visualize the treatment target and OARs with co-registered multiple contrast weightings. This conclusion is also backed by the inter-observer agreement on the contours of the tumor target and select OARs where our mean Dice coefficients and HD95 are better than those reported by recent literature (39,40). Among the five patients, the motion ranges of the tumor center-of-mass derived from MT images are within 2 pixels in all directions (SI, AP, LR) compared to those derived from StarVIBE images, expect for patient 4, where the StarVIBE images appeared to have failed in resolving respiratory motion. For these patient cases, the conventional protocol (a combination of T1w StarVIBE and the T2 BLADE sequences) required 12 min to provide adequate visualization of the tumor and OARs and depiction of motion, with the T2w contrast clearly revealing additional details of intra-tumoral structures. However, for the reference T2w sequence based on a 2D acquisition with thicker slice thickness, the visualization of the tumor was compromised when it was reformatted into coronal planes. Our proposed MT sequence requires only 8 min and offered largely uncompromised 3D reformatting for the T2w contrast in addition to respiratory motion resolution and other contrast weightings.

There are several limitations with this work. Firstly, the presented technique is optimized for 3T. However, 1.5T MR systems and, in some institutions, even low-field MR-guided radiotherapy systems are being used for abdominal RTP. Therefore, the technique may need to be further optimized. Secondly, because of the use of water excitation RF pulses, no fat information is available. Separate water and fat images from the same scan are presumably useful for better target and OARs delineation in the abdomen. It is therefore desired in the future to incorporate Dixon echoes in the pulse sequence to further enhance the capability of this proposed framework. Lastly, the current reconstruction framework, implemented in MATLAB, takes approximately 10 minutes on a workstation with moderate hardware configuration, which is sufficient for RTP workflow, in which the MR imaging session occurs days before treatment. However, this reconstruction time is not yet ideal for potential application in online real-time MR-guided radiation therapy, which needs to occur minutes before the treatment session. To further increase reconstruction speed, more efficient programming languages such as C++ and deep learning accelerated reconstruction can be explored.

5. Conclusion

In this work, we developed an MR Multitasking-based, abdomen-dedicated technique capable of providing motion-resolved, multi-contrast volumetric images in a single scan. These technical advantages may translate into improved accuracy, precision, and efficiency in abdominal radiotherapy planning. Further validation in a clinical setting is warranted.

Supplementary Material

Supinfo

[Supporting Information Figure S1]

Comparison between MT images before (A) and after (B) enhancement by deformable vector fields (DVFs).

Video S1

[Supporting Information Video S1]

Demonstration of the binning mechanism. A frame of real-time image is first processed with image open. Then the liver dome can be tracked across multiple contrasts through thresholding.

Video S2

[Supporting Information Video S2]

Illustration of the motion range comparison between MT images at T1 contrast (right) and the 2D real-time reference (left). Motion range was measured at the liver dome.

Video S3

[Supporting Information Video S3]

Video demonstration of the MT images across contrasts, motion states and slices from two healthy volunteers. The left figure is the same volunteer demonstrated in Figure 5.

Video S4

[Supporting Information Video S4]

Video demonstration of the retrospective real-time multi-contrast images, with both changing contrast and contrast-frozen images.