Multi-contrast four-dimensional magnetic resonance imaging (MC-4D-MRI): Development and initial evaluation in liver tumor patients

Abstract

Purpose:

To develop a novel multi-contrast four-dimensional magnetic resonance imaging (MC-4D-MRI) technique that expands single image contrast 4D-MRI to a spectrum of native and synthetic image contrasts and to evaluate its feasibility in liver tumor patients.

Methods and materials:

The MC-4D-MRI technique integrates multi-parametric MRI fusion, 4D-MRI, and deformable image registration (DIR) techniques. The fusion technique consists of native MRI as input, image preprocessing, fusion algorithm, adaptation, and fused multi-contrast MRI as output. Four-dimensional deformation vector fields (4D-DVF) were generated from an original T2/T1-w 4D-MRI by deforming end-of-inhalation (EOI) to nine other phase volumes via DIR. The 4D-DVF were applied to multi-contrast MRI to generate a spectrum of 4D-MRI in different image contrasts. The MC-4D-MRI technique was evaluated in five liver tumor patients on tumor contrast-to-noise ratio (CNR), internal target volume (ITV) contouring consistency, diaphragm motion range, and tumor motion trajectory; and in digital anthropomorphic phantoms on 4D-DIR introduced errors in tumor motion range, centroid location, extent, and volume.

Results:

MC-4D-MRI consisting of 4D-MRIs in native image contrasts (T1-w, T2-w, and T2/T1-w) and synthetic image contrasts, such as tumor-enhanced contrast (TEC) were generated in five liver tumor patients. Patient tumor CNR increased from 2.6 ± 1.8 in the T2/T1-w MRI, to −4.4 ± 2.4, 6.6 ± 3.0, and 9.6 ± 3.9 in the T1-w, T2-w, and TEC MRI, respectively. Patient ITV inter-observer mean Dice similarity coefficient (mDSC) increased from 0.65 ± 0.10 in the original T2/T1-w 4D-MRI, to 0.76 ± 0.14, 0.77 ± 0.12, and 0.86 ± 0.05 in the T1-w, T2-w, and TEC 4D-MRI, respectively. Patient diaphragm motion range absolute differences between the three new 4D-MRIs and original T2/T1-w 4D-MRI were 1.2 ± 1.3, 0.3 ± 0.7, and 0.5 ± 0.5 mm, respectively. Patient tumor displacement phase-averaged absolute differences between the three 4D-MRIs and the original 4D-MRI were 0.72 ± 0.33, 0.62 ± 0.54, and 0.74 ± 0.43 mm in the superior-inferior (SI) direction, and 0.59 ± 0.36, 0.51 ± 0.30, and 0.50 ± 0.24 mm in the anterior-posterior (AP) direction, respectively. In the digital phantoms, phase-averaged absolute tumor centroid shift caused by the 4D-DIR were at or below 0.5 mm in SI, AP, and left-right (LR) directions.

Conclusion:

We developed an MC-4D-MRI technique capable of expanding single image contrast 4D-MRI along a new dimension of image contrast. Initial evaluations in liver tumor patients showed enhancements in image contrast variety, tumor contrast, and ITV contouring consistencies using MC-4D-MRI. The technique might offer new perspectives on the image contrast of MRI and 4D-MRI in MR-guided radiotherapy.

1 ∣. INTRODUCTION

The application of magnetic resonance imaging (MRI) in radiotherapy (RT) has been rising in the past decade, and might be further accelerated with the increasing installation of MRI simulation and hybrid MRI-RT machines in radiation oncology departments.^1-6 One of the major advances of MR-guided RT (MRgRT) focuses on abdominal and mobile tumors, where MRI shows advantages in its non-ionization nature, superior soft-tissue contrast, and motion imaging capability. Analog to four-dimensional CT (4D-CT) techniques in managing mobile tumors,⁷ various four-dimensional MRI (4D-MRI) techniques have been developed in the management of mobile lung and abdominal tumors, mostly within the past decade.^8-17

Despite many advances, current 4D-MRI techniques still face various known challenges, such as extended image acquisition time,⁸ limited temporal or spatial resolutions,^13,18,19 and artifacts caused by breathing variations.^20,21 However, a much less noticed aspect is image contrast, which is widely known as a unique advantage of MRI. Current 4D-MRIs are typically reconstructed from single-sequence or image contrast MRI, such as T1-w, T2-w, T2/T1-w and diffusion-weighted MRI (DWI).¹⁷ Different degrees of variation in tumor delineation or measurement have been observed from single image contrast MRIs.^22-25 These variations could be attributed to intrinsic inter-patient variation in tumor contrast, and/or sub-optimal tumor contrast of certain MRI sequences. The improvement of MRI tumor contrast and image contrast variety might be helpful for MRgRT applications, such as tumor delineation,^26-28 motion management,²⁹ image registration,^30,31 dose calculation,^32,33 among others.

A number of attempts have been made in the studying of MRI tumor contrast, clinical consistencies, and image contrast variety of 3D- and 4D-MRIs. For example, Zhang et al.²⁴ reported lower inter-observer variation in gross tumor volume (GTV) contouring on higher tumor contrast-to-noise ratio (CNR) MRI. Lee and Riederer³⁴ showed that improved tumor CNR can be achieved by linear-weighted summation of multiple native MRIs. Recently, Freedman et al.³⁵ demonstrated the feasibility of synthesizing T2-w 4D-MRI from T1-w 4D-MRI in lung cancer patients. In their study, 4D motion from the original T1-w 4D-MRI was transferred to MRI in another image contrast. The application of 4D motion model to multiple MRI sets in a variety of image contrasts is yet to be investigated. Furthermore, multiple native MRIs are often acquired for a patient while might be utilized separately. How to integrate and synergize the complementary information of multi-parametric MRI is yet to be explored.

In this study, we aimed to: (a) integrate multi-parametric MRI for a spectrum of MRIs in different image contrasts, including native image contrasts and new synthetic image contrasts and (b) transfer 4D motion model to them to generate a variety of 4D-MRIs in diverse image contrasts. We hypothesized that this multi-contrast 4D-MRI (MC-4D-MRI) technique could provide one approach in enhancing the image contrast variety, tumor contrast, and delineation consistencies of 4D-MRI.

This MC-4D-MRI technique is composed of three techniques: multi-parametric MRI fusion for multi-contrast generation, 4D-MRI, and deformable image registration (DIR) for patient-specific motion modeling. The MRI fusion technique was previously developed.³⁶ The 4D-MRI and DIR techniques are open to different 4D-MRI sorting methods,^16,20,37,38 motion modeling methods,^39-42 and DIR algorithms.^43-45 We hypothesized that, by combining the three techniques, a spectrum of 4D-MRIs in various image contrasts can be generated simultaneously. Compared to current 4D-MRI techniques, MC-4D-MRI is featured with a new dimension of “image contrast”, which is independent of the respiratory motion. In this sense, MC-4D-MRI could also be identified as 5D-MRI wherein the fourth dimension is the respiratory motion and the fifth dimension is the image contrast. The feasibility of the MC-4D-MRI technique was examined in five liver tumor patients and digital anthropomorphic phantoms.

2 ∣. MATERIALS AND METHODS

The workflow of the MC-4D-MRI technique is shown in Figure 1. Three major technical components are multi-parametric MRI fusion, 4D-MRI, and DIR.

FIGURE 1.

Workflow of the multi-contrast four-dimensional magnetic resonance imaging (MC-4D-MRI) technique. Left box: multi-parametric MRI fusion for multi-contrast MRI generation. Right box: generation of the MC-4D-MRI by integrating the major components, including original 4D-MRI, three-dimensional deformable image registration (3D-DIR) for alignment, and 4D-DIR for four-dimensional deformation vector fields (4D-DVF) generation

2.1 ∣. Multi-parametric MRI fusion

A multi-parametric MRI fusion technique previously developed was used to synthesize multi-contrast MRI (MC-MRI). Detailed description of the method can be found in the previous report³⁶ and only briefly summarized here. As shown in the left box of Figure 1, the fusion consists of input multi-parametric MRI, image preprocessing, fusion algorithm, adaptation, and fused MC-MRI. Input multi-parametric MR images were native MRI acquired under end-of-inhalation (EOI) phase in breath-hold. All input MRI (e.g., T1-w, T2-w) were co-registered to the reference (T2/T1-w, EOI) by DIR function in MIM Maestro v6.0 (MIM Software Inc., Cleveland, OH, USA). Image intensities were clipped to the 99.5th percentile of each image set and normalized to values between 0 and 1.

Linear-weighted summation algorithm was used for the fusion. The MC-MRI were synthesized by the equation:

$$Y_i=\sum _{k=1}^K{w}_{ik}{X}_k,$$ (1)

where X_k are input MRI and w_ik ∈ [−1, 1] in an interval of 0.2 are the weight coefficients of the kth input MRI in the ith fused MRI.

A database of MC-MRI with weights and image features was built for each patient. In this study, the features of interest were tumor CNR and liver signal-to-noise ratio (SNR):

$${\text{CNR}}_{\text{tumor}}=\frac{\mu (\text{tumor})-\mu (\text{liver})}{\sigma (\text{liver})},$$ (2)

$${\text{SNR}}_{\text{liver}}=\frac{\mu (\text{liver})}{\sigma (\text{liver})},$$ (3)

where liver region was a homogenous liver area near the tumor. MC-MRI were generated by two methods: weight-driven by interactively tuning the weights of each input MRI, and feature-driven by constraining the tumor CNR and liver SNR of the fused image. Tumor-enhanced contrast (TEC) MRI was generated by the feature-driven method, where the constraints were set as maximizing tumor CNR and a positive liver SNR.

2.2 ∣. 4D-MRI technique

There is no restriction on the type of sequence or sorting method for the original 4D-MRI. In this study, T2/T1-w Fast Imaging Employing Steady-state Acquisition (FIESTA) sequence on a GE Signa HDx 1.5T scanner in 2D cine mode was used. Each 2D slice was imaged for 12 s at 0.3 s/frame. Imaging parameters were: time of repetition (TR), 3.7 ms; time of echo (TE), 1.21 ms; flip angle, 52°; slice thickness, 5 mm; in-plane resolution, 1.88 mm × 1.88 mm; bandwidth (BW), 977 Hz/pixel. Raw images were retrospectively sorted to 10 phase bins using body area as internal surrogate as previously reported.^10,46 The reconstructed T2/T1-w 4D-MRI was defined as the original 4D-MRI in this study.

2.3 ∣. 4D-DIR, 3D-DIR, and MC-4D-MRI

EOI phase of the original 4D-MRI was defined as the reference (MRI_ref). It was deformed to the other nine phases by DIR in MIM Maestro software to generate nine deformation vector fields (4D-DVF). This process was defined as 4D-DIR. If the reference MRI in the fusion process (Section 2.1) was in different phase or contrast from this MRI_ref, the MC-MRI would be deformed to MRI_ref for geometric realignment by 3D-DIR. In this study, no realignment was needed since the reference MRI in the fusion process was the same as MRI_ref.

Each image set of the MC-MRI was deformed by the 4D-DVF to render nine volumes, forming the other nine phase volumes of a new 4D-MRI. Simultaneously, MC-MRI in native or synthetic image contrasts were expanded to multiple 4D-MRI sets. This spectrum of 4D-MRIs in diverse image contrast was defined as the MC-4D-MRI.

2.4 ∣. Digital anthropomorphic phantom XCAT and liver tumor patient study

Five sets of extended cardiac torso (XCAT) digital anthropomorphic phantoms were constructed to evaluate the errors introduced by the 4D-DIR method. Patient T2/T1-w imaging parameters were used for the simulation. Spatial resolution was 1.88 mm × 1.88 mm × 5 mm. Patient-specific motion patterns and image intensities of the major organs and tumor were obtained from actual FIESTA T2/T1-w images in liver tumor patients. Hypothetic spherical tumors in 30 mm diameter were inserted in different locations of the liver to simulate real patient cases.

Five cancer patients with primary or metastatic liver tumors were included in this study. All MRI scans were performed on a GE Signa HDx 1.5T MR scanner. Original 4D-MRI was acquired and reconstructed as described in Section 2.2. T1-w Fast Spoiled Gradient Recalled Acquisition in the Steady State (FSPGR) and T2-w Fast Recovery Fast Spin Echo (FRFSE) MRI were acquired in axial planes at EOI phase during breath-hold. Imaging parameters were: in-plane resolution, 1.88 mm × 1.88 mm; slice thickness, 3 mm; BW, 391 Hz/pixel; TR, 225 ms (T1-w) and 2000 ms (T2-w); TE, 4.37 ms (T1-w) and 87.5 ms (T2-w). The study protocol was approved by Institutional Review Board and informed consents were obtained.

2.5 ∣. Evaluation of MC-4D-MRI

Image contrasts of the MC-MRI were evaluated in patients qualitatively on multi-contrast generation and quantitatively on tumor CNR and its inter-patient consistency. Tumor CNR were compared between original T2/T1-w, native (T1-w, T2-w), and synthetic TEC MRI. Mean, standard deviation (SD), and coefficient of variance (CoV) of tumor CNR across patients were evaluated for each image contrast, where CoV was defined as

$$\text{CoV}(\text{CNR})=\frac{\sigma ({\text{CNR}}_{\mathrm{i}})}{\mu ({\text{CNR}}_{\mathrm{i}})}\times 100\%,$$ (4)

where i is the index of patient. Wilcoxon signed-rank test was used for inter-contrast comparisons between original T2/T1-w and the other three MC-MRI contrasts. Bonferroni correction was applied to adjust for multiple comparison effect.

Internal target volume (ITV) contouring consistencies were evaluated by inter-observer contour mean Dice similarity coefficient (mDSC) and inter-patient variation of the mDSC. Maximum intensity projection (MIP) of the T2/T1-w, T2-w, and TEC4D-MRI and minimum intensity projection (MinIP) of the T1-w 4D-MRI were exported to Eclipse (Varian Medical Systems, Palo Alto, CA, USA) for ITV delineation. Three radiation oncologists performed the contouring independently. DSC was calculated between each pair of observer contours,

$$\text{DSC}=\frac{2(\mid {\text{ITV}}_{\mathrm{A}}\mid \cap \mid {\text{ITV}}_{\mathrm{B}}\mid )}{\mid {\text{ITV}}_{\mathrm{A}}\mid +\mid {\text{ITV}}_{\mathrm{B}}\mid },$$ (5)

where A and B are any combination of two observers. The mean DSC of all observer pairs was used to determine the inter-observer ITV contouring consistency and was defined as mDSC. Mean, SD, and CoV of the mDSC across patients were calculated for each image contrast, where CoV was defined as

$$\text{CoV}(\text{mDSC})=\frac{\sigma ({\text{mDSC}}_i)}{\mu ({\text{mDSC}}_i)}\times 100\%,$$ (6)

where i is the index of patient. Geometric and motion errors introduced by 4D-DIR were evaluated on patient-specific XCAT phantoms. Tumor centroid position, extent, volume, and motion patterns were measured after tumor segmentation by an in-house developed threshold-based method. The metrices were compared between original T2/T1-w 4D-MRI and its 4D-DIR warped 4D-MRI. The 4D-DIR warped 4D-MRI was generated by applying 4D-DVF to MRI_ref, in the same manner as the MC-4D-MRI generation. The differences between the two 4D-MRIs were therefore solely introduced by the 4D-DIR method. The same 4D-DIR warping method was used in the patient cases to evaluate the 4D-DIR introduced errors in the patients.

In patient cases, structural similarity indices (SSIM)⁴⁷ were measured between the original 4D-MRI and its 4D-DIR warped 4D-MRI. Luminance, contrast, and structure sub-indices were calculated as:

$$\begin{gathered} I(x,y)=\frac{2{\mu }_x{\mu }_y+C_1}{{\mu }_x^2+{\mu }_y^2+C_1},c(x,y)=\frac{2{\sigma }_x{\sigma }_y+C_2}{{\sigma }_x^2+{\sigma }_y^2+C_2}, \\ \text{and}s(x,y)=\frac{{\sigma }_{xy}+C_3}{{\sigma }_x{\sigma }_y+C_3}, \end{gathered}$$ (7)

where μ_x, μ_y, σ_x, σ_y, σ_xy are the local mean, SD, and cross-covariance of the two images in comparison. SSIM were calculated as:

$$\text{SSIM}=I(x,y)\cdot c(x,y)\cdot s(x,y),$$ (8)

with C₁ = (k₁L)², C₂ = (k₂L)², and C₃ = C₂ /2, where L = 255, k₁ = 0.01, and k₂ = 0.03.

Diaphragm motion range and tumor motion trajectory of the five 4D-MRIs (original, warped T2/T1-w, T1-w, T2-w, and TEC) were measured using a region-matching based method.⁴⁸ Diaphragm motion range were compared between original 4D-MRI and other four 4D-MRIs. Tumor motion trajectories were compared phase-by-phase between original 4D-MRI and other four 4D-MRIs, in SI and AP directions. The cross-phase mean and SD of the absolute differences were evaluated.

3 ∣. RESULTS

3.1 ∣. Multi-parametric MRI fusion

Figure 2 shows the multi-parametric MRI fusion result of a liver tumor patient. From three native-contrast MRIs (T2/T1-w, T1-w, and T2-w), a variety of synthetic-contrast MRIs were generated (nine are shown). Tumor was barely visible in the T2/T1-w (a), hypointense in the T1-w (b), and hyperintense in the T2-w (c) MRI. T2/T1-w MRI features high aorta signal, while T1-w MRI features high liver and muscle signal. By feature-driven method, TEC MRI which has the maximum tumor CNR and a positive liver SNR was synthesized automatically (e). By weight-driven method, hyperintense tumors accompanying organs such as liver, spleen, and aorta in different degrees of contrasts were generated (d–i, k). Other features like liver enhancement (j) and homogenous soft-tissue (l) were also generated by weight-driven method.

FIGURE 2.

Multi-parametric magnetic resonance imaging (MRI) fusion of a liver tumor patient. (a–c): Input native-contrast MRIs. (d–l): Synthetic-contrast MRIs. Thick arrow: tumor, thin arrow: aorta. A variety of new image contrasts were generated, including tumor-enhanced (e), hyperintense tumor with different combinations of tumor, liver, aorta, spleen (d–i, k), liver-enhanced (j), and others (l)

3.2 ∣. Tumor CNR of MC-MRI

Figure 3 shows box-plot of tumor CNR in four representative contrasts of the MC-MRI in liver tumor patients. Tumor CNR increased from 2.6 ± 1.8 in the T2/T1-w MRI, to −4.4 ± 2.4, 6.6 ± 3.0, and 9.6 ± 3.9 in T1-w, T2-w, and TEC MRI, respectively (p = 0.007, 0.032, and 0.016). The CoV of tumor CNR decreased from 69.5% (T2/T1-w), to 54.2% (T1-w), 46.3% (T2-w), and 40.9% (TEC). The median and minimum tumor CNR among the patients were 2.2 and 1.1 in T2/T1-w MRI, while they were 6.5 and 3.3 in T2-w, and 9.2 and 5.1 in the TEC MRI.

FIGURE 3.

Tumor contrast-to-noise ratio (CNR) of four contrasts in the multi-contrast magnetic resonance imaging (MC-MRI) of liver tumor patients. Original MRI (T2/T1-w) is shown on the left, native-contrast (T1-w, T2-w) MRI in the middle, and synthetic tumor-enhanced contrast (TEC) MRI on the right. Absolute values of T1-w tumor CNR were shown in the plot for easy visual comparison

3.3 ∣. MC-4D-MRI generation

Figure 4 demonstrates MC-4D-MRI of an example liver tumor patient. From one original 4D-MRI (T2/T1-w) and two native MRI volumes (T2-w and T1-w), a variety of 4D-MRI in native contrasts (T2-w and T1-w) and synthetic contrasts (C3-4, C6-7) were generated. The contrast features included high tumor contrast (C2-3), balanced tumor and soft-tissue signal (C4), high liver signal and contrast (C5-6), and homogenous soft-tissue signal (C7), among others. MC-4D-MRI were successfully generated for all liver tumor patients.

FIGURE 4.

Multi-contrast four-dimensional magnetic resonance imaging (MC-4D-MRI) of a liver tumor patient. Each row is a unique 4D-MRI (C1–7) in 10 respiratory phases (P1–10). C1: original T2/T1-w 4D-MRI, C2: T2-w 4D-MRI, C5: T1-w 4D-MRI, C3: synthetic tumor-enhanced contrast (TEC) 4D-MRI, and C4, 6, 7: other synthetic-contrast 4D-MRIs

The diversity of contrasts in MC-4D-MRI may provide unique and complementary information for different MRgRT applications. For instance, when one type of 4D-MRI couldn’t provide satisfactory tumor contrast (e.g., T2/T1-w, C1), other native-contrast (T2-w, C2) or synthetic-contrast (TEC, C3) 4D-MRIs in higher tumor contrast might be alternatives. Images like liver-enhanced 4D-MRI (C6) and homogeneous soft-tissue 4D-MRI (C7) might facilitate 4D normal-tissue segmentations or bulk-density-assignment based synthetic-CT generation.

3.4 ∣. Evaluation of ITV contour consistency of MC-4D-MRI

Figure 5 summarizes inter-observer and inter-patient ITV contouring consistencies of the MC-4D-MRI in the patients. The inter-observer mDSC were 0.65 ± 0.10, 0.76 ± 0.14, and 0.77 ± 0.12 for the T2/T1-w, T1-w, and T2-w 4D-MRI, respectively. TEC 4D-MRI showed the highest mDSC at 0.86 ± 0.05 (p = 0.008). This trend of improvement agrees with the tumor CNR increase from T2/T1-w MRI to T1-w, T2-w, and TEC MRI (Section 3.2). Minimum mDSC among the five patients were 0.51, 0.52, and 0.62 in T2/T1-w, T1-w, and T2-w 4D-MRI, respectively, while was 0.79 in TEC 4D-MRI. The CoV of mDSC across patients decreased from 15.4%, 18.4%, 15.6% (T2/T1-w, T1-w, T2-w 4D-MRI) to 5.8% in the TEC 4D-MRI.

FIGURE 5.

Internal target volume (ITV) contouring inter-observer mean Dice similarity coefficient (mDSC) of four contrasts in the multi-contrast four-dimensional magnetic resonance imaging (MC-4D-MRI) of liver tumor patients. Three native-contrast (T2/T1-w, T1-w, T2-w) and one synthetic-contrast (tumor-enhanced contrast, TEC) 4D-MRI were evaluated. Each box represents one image contrast

3.5 ∣. Estimation of 4D-DIR introduced errors by XCAT simulation

Patient-specific XCAT phantoms were simulated to estimate the potential errors introduced by the 4D-DIR method. Figure 6 shows EOI phase of the T2/T1-w 4D-XCAT for the five patients. Tables 1-4 summarize the tumor motion range, centroid position, extent, and volume differences between 4D-DIR warped 4D-MRI and the original 4D-MRI on the XCAT phantoms.

FIGURE 6.

Extended cardiac torso (XCAT) phantoms in real patient image resolution, tumor location, and motion pattern. (a–e) End-of-inhalation (EOI) phase of five 4D-XCAT phantoms with patient-specific tumor location and motion patterns. Left to right: axial, coronal, and sagittal planes showing the tumor center. Tumor: orange color sphere

TABLE 1.

Tumor motion range comparison between four-dimensional deformable image registration (4D-DIR) warped and original four-dimensional magnetic resonance imaging (4D-MRI) in patient-specific extended cardiac torso (XCATs)

	Original T2/T1-w 4D-MRI (mm)			Warped T2/T1-w 4D-MRI (mm)			Absolute difference (mm)
XCAT	SI	AP	LR	SI	AP	LR	SI	AP	LR
1	10.9	5.1	0.5	10.5	4.5	0.5	0.4	0.6	0.0
2	11.7	4.9	0.3	11.6	5.2	0.3	0.2	0.3	0.0
3	7.9	3.9	0.2	8.6	4.0	0.7	0.8	0.1	0.5
4	8.7	4.1	0.2	9.2	4.5	0.3	0.5	0.4	0.1
5	10.3	2.9	0.2	11.0	3.0	0.5	0.7	0.1	0.2
Mean	9.9	4.2	0.3	10.2	4.2	0.5	0.5	0.3	0.2
SD	1.6	0.9	0.1	1.2	0.8	0.2	0.2	0.2	0.2

TABLE 4.

Normalized tumor volume phase level absolute differences between four-dimensional deformable image registration (4D-DIR) warped and original four-dimensional magnetic resonance imaging (4D-MRI) in patient-specific extended cardiac torsos XCATs

XCAT	Mean across 10 phases	SD across 10 phases
1	3.2%	2.8%
2	1.5%	1.2%
3	1.9%	1.1%
4	3.7%	2.9%
5	3.4%	2.2%
Mean	2.7%	2.0%
SD	1.0%	0.8%

Table 1 summarizes the tumor motion range and the absolute differences between the two 4D-MRIs. Tumor motion range of original 4D-MRI were 9.9 ± 1.6 and 4.2 ± 0.9 mm in the SI and AP directions, respectively. Tumor motion range in the 4D-DIR warped 4D-MRI were 10.2 ± 1.2 and 4.2 ± 0.8 mm in the two directions. The tumor motion range absolute differences between the two 4D-MRIs were 0.5 ± 0.2 and 0.3 ± 0.2 mm, in the SI and AP directions, respectively. Phase level comparison of tumor centroid position between the two 4D-MRIs is summarized in Table 2. For all five phantoms, the tumor centroid position cross-phase mean absolute differences were at or below 0.5 mm in the three orthogonal directions (SI, AP and LR). The cross-phase SD were all below 0.35 mm. Table 3 summarizes the phase level comparison of tumor extent in SI, AP and LR directions between the two 4D-MRIs. The cross-phase mean and SD of the tumor extent absolute differences were around or below 1 mm in the three orthogonal directions. Phase level comparison of tumor volume between 4D-DIR warped and original 4D-MRI is summarized in Table 4. The tumor volume cross-phase mean absolute differences were 2.7 ± 1.0% for the five phantoms. The cross-phase SD was 2.0 ± 0.8% for the five phantoms.

TABLE 2.

Tumor centroid position phase level absolute differences between four-dimensional deformable image registration (4D-DIR) warped and original four-dimensional magnetic resonance imaging (4D-MRI) in patient-specific extended cardiac torso (XCATs)

XCAT	Mean across 10 phases (mm)			SD across 10 phases (mm)
	SI	AP	LR	SI	AP	LR
1	0.22	0.13	0.17	0.19	0.19	0.15
2	0.31	0.17	0.15	0.30	0.14	0.09
3	0.29	0.13	0.18	0.24	0.10	0.17
4	0.50	0.15	0.09	0.29	0.17	0.06
5	0.38	0.19	0.12	0.33	0.13	0.09
Mean	0.34	0.15	0.14	0.27	0.14	0.11
SD	0.10	0.03	0.04	0.05	0.03	0.05

TABLE 3.

Tumor extent phase level absolute differences between four-dimensional deformable image registration (4D-DIR) warped and original four-dimensional magnetic resonance imaging (4D-MRI) in patient-specific extended cardiac torso (XCATs)

XCAT	Mean across 10 phases (mm)			SD across 10 phases (mm)
	SI	AP	LR	SI	AP	LR
1	0.8	0.6	0.4	0.63	0.52	0.52
2	0.9	0.4	0.9	0.88	0.52	0.57
3	0.8	0.9	1.1	0.79	0.99	0.74
4	1.6	0.3	0.7	1.17	0.48	0.48
5	1.0	0.8	0.2	0.94	0.63	0.42
Mean	1.02	0.60	0.66	0.88	0.63	0.55
SD	0.33	0.25	0.36	0.20	0.21	0.12

3.6 ∣. Estimation of 4D-DIR introduced errors in patients

Figure 7 illustrates the original and 4D-DIR warped 4D-MRI in EOI and end-of-exhalation (EOE) phases of an example patient. Low signals in the absolute difference maps demonstrated the structural agreement between the 4D-DIR warped and original 4D-MRI in the two boundary phases of a respiratory cycle. Similar patterns were observed in the other eight phase volumes.

FIGURE 7.

Visualization and comparison of four-dimensional deformable image registration (4D-DIR) warped and original four-dimensional magnetic resonance imaging (4D-MRI) of a liver tumor patient. (a and b) End-of-inhalation (EOI) and end-of-exhalation (EOE) phases of the original and warped 4D-MRI, and the absolute difference map in coronal view

Table 5 summarizes the phase level SSIM indices between the 4D-DIR warped and original 4D-MRI. The cross-phase mean SSIM and structure similarity sub-index were 0.841 ± 0.040 and 0.863 ± 0.038 for the five patients. The cross-phase SD of the two indices were 0.011 ± 0.004 and 0.009 ± 0.004 for the patients. The contrast and luminance similarity indices were both close to 1 for all patient cases.

TABLE 5.

Structural similarity (SSIM) indices between four-dimensional deformable image registration (4D-DIR) warped and original four-dimensional magnetic resonance imaging (4D-MRI) in liver tumor patients

Patient	Mean across 10 phases				SD across 10 phases
	SSIM	Structure	Contrast	Luminance	SSIM	Structure	Contrast	Luminance
1	0.888	0.904	0.985	0.988	0.006	0.004	0.001	0.003
2	0.864	0.883	0.983	0.988	0.012	0.010	0.002	0.002
3	0.826	0.854	0.984	0.975	0.009	0.007	0.001	0.003
4	0.847	0.868	0.982	0.988	0.013	0.010	0.002	0.005
5	0.782	0.804	0.971	0.984	0.017	0.015	0.004	0.006
Mean	0.841	0.863	0.981	0.985	0.011	0.009	0.002	0.004
SD	0.040	0.038	0.006	0.005	0.004	0.004	0.001	0.002

The 4D-DIR introduced errors in patients were further evaluated on the diaphragm motion range and tumor motion trajectories. As shown in the first of the four comparisons in Tables 6 and 7, the diaphragm motion range absolute differences between 4D-DIR warped and original 4D-MRI were 0.3 ± 0.4 mm for the five patients. The tumor displacement cross-phase mean absolute differences were 0.42 ± 0.16 and 0.44 ± 0.36 mm, in the SI and AP directions, respectively.

TABLE 6.

Diaphragm motion range comparison between original four-dimensional magnetic resonance imaging (4D-MRI) and multi-contrast 4D-MRI (MC-4D-MRI) in liver tumor patients

Patient	Diaphragm SI motion range (mm)					Absolute difference from Original (mm)
	Original	T2/T1-w	T1-w	T2-w	TEC	T2/T1-W	T1-w	T2-w	TEC
1	12.0	12.0	12.0	12.0	12.0	0	0	0	0
2	12.0	12.0	12.0	12.0	12.5	0	0	0	0.5
3	10.0	9.0	9.0	10.0	10.0	1.0	1.0	0	0
4	9.0	9.5	7.0	9.0	8.0	0.5	2.0	0	1.0
5	11.0	11.0	8.0	9.5	12.0	0	3.0	1.5	1.0
Mean	10.8	10.7	9.6	10.5	10.9	0.3	1.2	0.3	0.5
SD	1.3	1.4	2.3	1.4	1.9	0.4	1.3	0.7	0.5

Abbreviation: TEC, tumor-enhanced contrast.

TABLE 7.

Tumor motion trajectory phase level absolute differences between original four-dimensional magnetic resonance imaging (4D-MRI) and multi-contrast 4D-MRI (MC-4D-MRI) in liver tumor patients

SI Patient	Mean across 10 phases (mm)				SD across 10 phases (mm)
	T2/T1-W	T1-w	T2-w	TEC	T2/T1-W	T1-w	T2-w	TEC
1	0.25	0.25	0.25	0.35	0.35	0.42	0.26	0.34
2	0.30	0.60	0.35	0.30	0.35	0.57	0.34	0.26
3	0.65	1.00	1.55	1.10	0.58	1.25	1.26	0.94
4	0.45	0.70	0.35	0.70	0.44	0.92	0.41	1.21
5	0.45	1.05	0.58	1.25	0.83	1.67	0.66	1.81
Mean	0.42	0.72	0.62	0.74	0.51	0.97	0.59	0.91
SD	0.16	0.33	0.54	0.43	0.20	0.51	0.40	0.64
AP Patient	Mean across 10 phases (mm)				SD across 10 phases (mm)
	T2/T1-W	T1-w	T2-w	TEC	T2/T1-W	T1-w	T2-w	TEC
1	0.20	0.20	0.15	0.20	0.26	0.26	0.24	0.26
2	0.35	0.65	0.55	0.60	0.34	0.41	0.50	0.39
3	1.05	0.70	0.80	0.60	0.90	0.82	0.54	0.46
4	0.15	0.30	0.25	0.30	0.24	0.42	0.42	0.42
5	0.45	1.10	0.80	0.80	0.55	1.90	0.96	1.03
Mean	0.44	0.59	0.51	0.50	0.46	0.76	0.53	0.51
SD	0.36	0.36	0.30	0.24	0.27	0.67	0.26	0.30

Abbreviation: SI, superior-inferior; AP: anterior-posterior; SD: standard deviation; TEC, tumor-enhanced contrast.

3.7 ∣. Evaluation of MC-4D-MRI diaphragm and tumor motion in patients

Table 6 summarizes the diaphragm motion range and comparisons between original 4D-MRI and MC-4D-MRI. The diaphragm motion range was 10.8 ± 1.3 mm in the original 4D-MRI, and were 10.7 ± 1.4, 9.6 ± 2.3, 10.5 ± 1.4, and 10.9 ± 1.9 mm in the warped T2/T1-w, T1-w, T2-w, and TEC 4D-MRI, respectively. The diaphragm motion range absolute differences between the original 4D-MRI and MC-4D-MRI were 0.3 ± 0.4, ± 1.3, 0.3 ± 0.7, and 0.5 ± 0.5 mm, for the four 4D-MRIs, respectively. It is worth mentioning that the spatial resolution of the motion measurements was 0.5 mm. Therefore, the 0 mm reported in individual comparisons were limited by this resolution, while the patient average values should be less affected.

Figure 8 illustrates tumor motion trajectories of MC-4D-MRI of an example patient. Five 4D-MRIs were evaluated, including original, 4D-DIR warped T2/T1-w, T1-w, T2-w, and TEC 4D-MRI. For each 4D-MRI, phase 1 (EOI) was set as the tumor motion trajectory origin in both SI and AP directions. Phase level absolute differences in tumor displacement between original and the other four 4D-MRIs for all patients were summarized in Table 7.

FIGURE 8.

Multi-contrast four-dimensional magnetic resonance imaging (MC-4D-MRI) tumor motion trajectory in SI and AP directions of a liver tumor patient. Horizontal axis is the respiratory phase index. Vertical axis is the tumor displacement in the SI (a) and AP (b) directions. Each line represents one 4D-MRI in the MC-4D-MRI set

As shown in Table 7, in the SI direction, tumor displacement cross-phase mean absolute differences from the original 4D-MRI were 0.42 ± 0.16, 0.72 ± 0.33, 0.62 ± 0.54, and 0.74 ± 0.43 mm, for the four 4D-MRIs (warped T2/T1-w, T1-w, T2-w, and TEC), respectively. The cross-phase SD of the absolute differences were 0.51 ± 0.2, 0.97 ± 0.51, 0.59 ± 0.4, and 0.91 ± 0.64 mm for the four 4D-MRIs, respectively. Similar levels of differences were observed in the AP direction, where tumor displacement cross-phase mean absolute differences from the original 4D-MRI were 0.44 ± 0.36, 0.59 ± 0.36, 0.51 ± 0.30, and 0.50 ± 0.24 mm, for the four 4D-MRIs, respectively. The cross-phase SD of the absolute differences were 0.46 ± 0.27, 0.76 ± 0.67, 0.53 ± 0.26, and 0.51 ± 0.30 mm for the four 4D-MRIs, respectively.

4 ∣. DISCUSSION

4.1 ∣. Methodology

In this study, a novel MC-4D-MRI technique capable of producing a variety of 4D-MRIs in native and synthetic MRI image contrasts was developed. By combining multi-parametric MRI fusion, 4D-MRI and DIR, the MC-4D-MRI technique synergized two features of MRI, image contrast diversity and motion imaging capability. In particular, the MRI fusion utilized and expanded the MRI diversity along the “image contrast” dimension. By linear-weighted summation or subtraction between multi-parametric MRIs, anatomical structures previously featured in single MRI were dynamically combined or enhanced in MC-MRI and MC-4D-MRI. The spectrum of image contrast is independent of the patient motion. Therefore, MC-4D-MRI could also be identified as 5D-MRI wherein the fourth dimension is respiratory motion and the fifth dimension is image contrast.

4.2 ∣. Clinical impact

Current 4D-MRI techniques generally rely on single MRI sequence or image contrast. Each of the native-contrast 4D-MRI has its own strength and weakness. For instance, T2/T1-w and T1-w 4D-MRI are relatively fast to acquire, while have sub-optimal tumor contrasts. T2-w and DWI 4D-MRI have higher tumor contrast, while require longer acquisition time. To acquire multiple 4D-MRIs can be more time consuming. The proposed MC-4D-MRI technique examined an approach of generating a large number of 4D-MRI sets, using an acquisition time of one relatively fast 4D-MRI and one set of multi-parametric MRI.

The increase in patient 4D-MRI set size and image contrast variety might provide unique or complementary information for different MRgRT applications. One potential clinical benefit was evaluated on the synthetic TEC 4D-MRI. Compared to the three native-contrast MRI, TEC MRI showed enhanced tumor CNR and inter-patient consistency. This might have contributed to the higher ITV delineation inter-observer and inter-patient consistencies on TEC 4D-MRI. The consistencies of ITV delineation play important roles in the quality of MRgRT treatment planning and dose delivery. TEC is one of the many image contrasts in MC-4D-MRI. We hope the added dimension of image contrast could contribute to the method developments in different aspects of MRgRT, such as target delineation, normal-tissue segmentation, bulk-density-assignment based synthetic-CT, among others.

The availability of multiple 4D-MRI sets can also be valuable. Among the common native contrasts, T1-w 4D-MRI has high liver and muscle signal, T2-w 4D-MRI has favorable hyperintense tumor contrast, while T2/T1-w 4D-MRI has high blood vessel signal. Having MC-4D-MRI available simultaneously might provide a dynamic 4D-MRI toolbox for the physicists and physicians, offering versatile information and utilities.

One clinical challenge less noticed in MRgRT is the intra-sequence inter-patient variation in tumor contrast. Single-sequence 4D-MRI might not always perform optimally for its intended purpose for all patients. When one set of 4D-MRI couldn’t provide the expected information, a variety of other 4D-MRIs might serve as alternatives. For instance, when T2-w 4D-MRI couldn’t provide satisfactory tumor contrast for certain patients, alternative TEC, T1-w, or T2/T1-w 4D-MRI in the “MC-4D-MRI tool-box” might be able to provide supplementing information, and vice versa.

4.3 ∣. Limitations

One limitation of the current MC-4D-MRI technique is the reliance on 4D motion computation. The accuracy of 4D-DIR and warping are critical for the quality of the MC-4D-MRI. In this study the potential errors introduced by 4D-DIR were evaluated on the original 4D-MRI in both patient-specific digital phantoms and patients. As summarized in Tables 1-7, the tumor centroid location absolute differences between 4D-DIR warped and original 4D-MRI in all three directions on the digital phantoms were at sub-millimeter level. In the patients, the SSIM and its sub-indices suggested reasonable structural agreement between original 4D-MRI and its 4D-DIR warped images. The tumor displacement cross-phase mean absolute differences were also at sub-millimeter level. These results were reassuring for the 4D-DIR method used in this study. Nevertheless, the sample size is relatively small. For larger clinical studies or practices, if high errors occurred in the 4D-DIR process, the quality of MC-4D-MRI and the clinical tasks could be adversely affected, such as under or over-estimation of ITV Therefore, it is important and necessary to perform quality assurance on the 4D-DIR introduced errors before applying 4D-DVF to the MC-4D-MRI generation.

Another limitation is the requirement of registration between input MRI and the original 4D-MRI reference phase. In its current form, the MC-4D-MRI technique generates one set of 4D-DVF from the original 4D-MRI and applies it to different image contrasts. Therefore, the quality of initial registration is important and would affect the quality of MC-4D-MRI. In this study, we used MIM Maestro DIR function for the 3D registration and inherited its registration accuracy. In Section 3.7, the MC-4D-MRI patient motion evaluation results included the errors introduced by both 3D-DIR registration and 4D-DIR.

It is worth noting that due to limited tumor contrast of certain 4D-MRI sets, consistent quantification of tumor absolute location was challenging. Alternatively, via a region-matching based method,⁴⁸ patient tumor and diaphragm motions were measured relative to the EOI of each 4D-MRI in this study. Therefore, Tables 6, 7 and Figure 8 only demonstrated the relative motion patterns of the MC-4D-MRI. They should be interpreted carefully to avoid over-estimation of the initial 3D-DIR registration accuracy, which was not separately evaluated in this study. Nevertheless, it is important to perform initial 3D registration in a high quality to achieve a high-quality MC-4D-MRI.

4.4 ∣. Future prospects

Recent developments in DIR and multi-parametric MRI methods have suggested potential approaches in improving the two limitations of the proposed MC-4D-MRI technique. Deep learning based DIR methods^49,50 have shown promises in the improvement of DIR accuracy. MR fingerprinting based methods^51,52 have indicated potentials for registration-free multi-parametric MRI acquisition methods.

Finally, a number of input MRI such as dynamic contrast-enhanced (DCE) MRI²² and DWI¹⁶ could be included in future multi-parametric MRI fusion to produce a greater range of image contrasts. It is worth noting that the term “contrast” refers to exogenous contrast agent in DCE-MRI, while refers to visual image contrast in this MC-4D-MRI study. In summary, with future advancements, the MC-4D-MRI technique could be further improved in its components for higher geometric accuracy and more diverse image contrasts.

5 ∣. CONCLUSION

A novel MC-4D-MRI or 5D-MRI technique that expands single image contrast 4D-MRI along a new dimension of image contrast was developed. Initial evaluations in liver tumor patients showed enhancements in image contrast variety, tumor contrast, and ITV contouring consistencies using MC-4D-MRI. The technique might offer new perspectives on the image contrast of MRI and 4D-MRI in MRgRT.

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.