Optical Flow-Guided Analysis of Intrafractional Anatomical Variations in Pancreatic Tumors and Organs at Risk During Magnetic Resonance-Guided Adaptive Radiation Therapy Under Abdominal Compression

Abstract

Purpose

This study aimed to investigate intrafractional anatomic variations in pancreatic tumors and organs at risk (OARs) during magnetic resonance-guided adaptive radiation therapy (MRgART) using optical flow calculations.

Methods and Materials

This study included 732,612 magnetic resonance (MR) slices from 345 fractions obtained from 25 consecutive patients with pancreatic cancer. All patients underwent MRgART using the Elekta Unity MR-Linac under abdominal compression. For each fraction, single-slice 2-dimensional cine MR images were acquired every 600 ms at the tumor center. On the first frames in the coronal and sagittal planes, the OARs (duodenum, stomach, small bowel, and colon) were delineated. Displacements were measured at 30-second intervals using the Farnebäck optical flow in the superior-inferior (SI) direction in the coronal and sagittal planes within the overlapping regions between each delineated structure and the rectangular region, with a side length of 10 cm centered on the tumor. To exclude outliers, range of motion in each direction was defined as the 2.5th–97.5th percentile of displacement, and statistical differences were assessed using the Wilcoxon signed-rank test with Holm-Bonferroni correction (P < .05).

Results

The median SI range of motion of the tumor was 5.8 mm in both the coronal and sagittal planes, which was significantly smaller than that of all OARs (P < .05). Among the OARs, the colon and small bowel (median, 8.9-10.2 mm) exhibited larger motion ranges than the duodenum and stomach (median, 6.3-7.6 mm; P < .05). The calculated margins in the SI direction were 3.1 to 3.4 mm for the tumor and 2.1 to 2.9 mm for the OARs.

Conclusions

Optical flow-guided analysis of intrafractional variations revealed that pancreatic tumors exhibited smaller ranges of motion than the adjacent OARs. These findings warrant direction-specific margins that account for intrafractional anatomic variations in both the tumors and surrounding OARs during pancreatic MRgART under abdominal compression.

Introduction

Pancreatic cancer (PC) has a poor prognosis and is predominantly diagnosed at an advanced stage.¹ Radiation therapy plays a crucial role in the local disease control, particularly for patients with unresectable or locally advanced PC.^2,3 However, the close proximity of highly radiosensitive organs at risk (OARs), such as the duodenum, stomach, small bowel, and colon, poses a substantial challenge for the safe delivery of high-dose radiation to tumors.⁴ Therefore, optimizing the balance between achieving adequate tumor dose coverage and minimizing doses to adjacent OARs remains a major clinical concern.

The accurate localization of abdominal radiation therapy requires the careful management of uncertainties that accrue from intrafractional and interfractional anatomic variations.⁵ Intrafractional variations occur during a single treatment session owing to factors such as respiration, peristalsis, or patient motion. In contrast, interfractional variations originate from day-to-day changes in patient anatomy, including differences in the size, shape, and location of the tumor and OARs. Although the internal and setup margins for both the planning target volume (PTV) and planning risk volume (PRV) compensate for these uncertainties, they should be minimized to prevent tumor underdosage and excessive doses to the adjacent OARs.⁶

Intrafractional anatomic changes during treatment, particularly respiration-induced organ motion, remain critical challenges in delivering precise radiation doses to abdominal tumors. In radiation therapy for abdominal organs, various motion-management techniques, such as abdominal compression (AC),7, 8, 9 breath-holding,¹⁰ and respiratory gating,¹¹ have been employed to minimize the intrafractional variations of the tumor and OARs. However, these techniques are inherently unable to address interfractional anatomic changes that require a distinct approach.

In recent years, adaptive radiation therapy (ART) has modified the daily treatment plans and thereby emerged as a transformative approach for addressing interfractional anatomic variations.¹² The clinical feasibility and dosimetric advantages of cone beam computed tomography (CBCT)-guided^13,14 and magnetic resonance (MR)-guided ART (MRgART) have been demonstrated.^15,16 Compared with CBCT images, MR provides superior soft-tissue contrast, which is particularly beneficial for the treatment of abdominal tumors.¹⁷ Furthermore, using 2-dimensional (2D) cine MR, MRgART enables the real-time monitoring of tumors and OARs during beam delivery without additional radiation exposure.^15,16 Besides 2D image acquisition, the analysis of abdominal cross-sections with high contrast and temporal resolution is essential for gaining novel insights into intrafractional anatomic variations in pancreatic MRgART.

To characterize the abdominal tumor motion during MRgART, 2D cine MR images have been employed.^9,18,19 Grimbergen et al⁹ analyzed the pancreatic tumor motion patterns during MRgART using an abdominal corset and emphasized the need for individualized motion assessments. Keiper et al¹⁸ and Jassar et al¹⁹ quantified the intrafractional motion of abdominal tumors using orthogonal 2D cine MR data with high temporal resolution. To analyze OAR motion, Wittmann et al²⁰ and Alam et al²¹ evaluated the intrafractional anatomic changes in the duodenum, stomach, colon, and small bowel using pretreatment and posttreatment 3-dimensional (3D) MR images and reported that such motion can have a dosimetric impact during MRgART. Damen et al²² developed a 2D tracking method for the small bowel that simultaneously optimized image acquisition and tracking accuracy, and enabled the real-time monitoring of OAR motion during beam delivery. However, these studies primarily addressed the intrafractional motion of either the tumor or OARs individually, without examining their spatiotemporal correlations.

This study investigated the intrafractional anatomic variations of both tumors and OARs and examined their spatiotemporal correlation using 2D cine MR images during MRgART under AC. We previously used an optical flow algorithm to quantify intrafractional liver motion and enable spatiotemporal motion analysis of arbitrary anatomic regions.⁷ Accordingly, we hypothesized that application of optical flow calculations to both tumors and OARs would enable the simultaneous evaluation of intrafractional motion and facilitate the investigation of how their motions are spatially and temporally correlated. The ranges of motion of the tumors and OARs were quantified in the left-right (LR), anterior-posterior (AP), and superior-inferior (SI) directions using optical flow calculations, and their correlations were investigated. This approach may provide critical insights into motion-management strategies and enable the determination of an appropriate margin design in pancreatic MRgART.

Methods and Materials

Patient cohort and procedure of radiation therapy

This study was performed in accordance with the Declaration of Helsinki and approved by the Medical Ethics Committee of the Graduate School of Medicine and Faculty of Medicine, Osaka Metropolitan University. It was conducted with permission from the head of the research institution (2021-273).

All patients provided informed consent for the use of their clinical data. A total of 732,612 MR slices were obtained from 345 cine MR series of 25 consecutive patients with PC. The patients underwent pancreatic MRgART using the Elekta Unity 1.5T MR-Linac (Elekta AB) under AC at Osaka Metropolitan University between September 2022 and March 2025. AC was selected in accordance with our institutional treatment protocol; this approach also helped maintain efficient treatment delivery during pancreatic MRgART. In radiation therapy planning, the gross tumor volume (GTV) was delineated by board-certified radiation oncologists. The clinical target volume (CTV) encompassed the GTV with a uniform 5-mm isotropic margin, the retropancreatic region, and the paraortic lymph node area between the celiac and superior mesenteric arteries.^23,24 The PTV was subsequently generated by expanding the CTV by an additional 3 to 5 mm margin in all directions.

Table E1 summarizes the patients’ characteristics. In brief, the participants included 14 men and 11 women with a median age of 67 years (range, 45-88 years). The numbers of patients with performance statuses of 0, 1, and 2 were 11, 13, and 1, respectively. According to the 8th edition of the Union for International Cancer Control criteria,²⁵ cT4N0M0 was the most common clinical stage, observed in 8 patients. Primary tumors were located in the pancreatic head in 16 patients and the body in 9 patients, with a median GTV of 27.4 cm³ (range, 2.9-164.7 cm³). Nineteen patients were treated with 42 to 54 Gy in 15 fractions, while 6 patients received 45 to 50 Gy in 10 fractions.

Acquisition of 2D cine MR images

The treatment workflow is shown in Figure E1. A 3D MR image for position verification was acquired immediately before irradiation, and treatment was initiated within 2 to 3 minutes of 3D MR image acquisition. Any discrepancy between the adaptive planning and treatment positions was corrected through adaptive replanning when necessary. Single-slice 2D cine MR images were acquired in the axial, coronal, and sagittal planes at the tumor center during irradiation, using a 2D balanced turbo field echo sequence. Each plane was imaged every 600 ms, with a 200-ms interval between consecutive planes. The scanning parameters were as follows: an echo time of 0.98 ms, a repetition time of 2.70 ms, an image resolution of 352 × 352 pixels, a pixel spacing of 1.2 mm, and a slice thickness of 5.0 mm.⁷ All captured 2D cine MR data were exported in binary format and converted to the portable network graphics format for further analysis.

Optical flow algorithm

Figure 1 shows the workflow of the optical flow calculations. First, the regions of interest (ROIs) of the tumors and OARs were defined in both the coronal and sagittal planes (Fig. 1a). For each fraction, a square with a side length of 10 cm that was centered on the tumor was created to represent the motion-monitoring ROI. PTV structures were used as tumor ROIs for motion monitoring in this study. OARs, including the duodenum, stomach, small bowel, and colon, were manually contoured on the first frame of both planes. The ROIs for the OARs were defined as the overlapping regions between each delineated structure and square regions. This area setting enabled the evaluation of intrafractional variations, specifically for clinically relevant OARs located close to the tumor. Next, the vectors of the tumors and OARs within the predefined ROIs were quantified in subregions in the horizontal and vertical directions at 4-pixel intervals using the Farnebäck optical flow algorithm in Python version 3.10.8 (Fig. 1b).^7,26 The Farnebäck algorithm computes dense motion fields by approximating the pixel brightness variation in 2 consecutive frames using a polynomial expansion.²⁶ Then, the displacement vector was decomposed into horizontal and vertical components every 600 ms in the coronal and sagittal planes (Fig. 1c). In the coronal plane, the horizontal and vertical directions were defined as the LR and SI (SI_coronal) directions, respectively. In the sagittal plane, they were defined as the AP and SI (SI_sagittal) directions, respectively. Finally, the calculated displacements were divided into 30-second segments, and the motion range was defined for each segment (Fig. 1d). The motion range in the LR, AP, SI_coronal, and SI_sagittal directions was defined as the 2.5th to 97.5th percentiles of the displacements to exclude 5% of the outliers.^7,27

Figure 1.

Workflow of the optical flow calculation, comprising 4 steps: (a) setting regions of interest of tumors and organs at risk (OARs) in the coronal and sagittal planes; (b) calculating 2-dimensional vectors using the Farnebäck optical flow algorithm; (c) decomposing the calculated vectors into vertical and horizontal axes; and (d) defining motion range every 30 seconds.

Abbreviations: AP = anterior-posterior; LR = left-right; PTV = planning target volume; SI = superior-inferior; SI_coronal = superior-inferior in the coronal plane; SI_sagittal = superior-inferior in the sagittal plane.

Analysis of baseline drift

As an analysis of intrafractional variation of the tumor and OARs, the temporal changes in the center of motion ranges (ie, baseline drift) were calculated. The position of origin of each structure was defined as the center of amplitude that was calculated from the first 30 seconds.⁹ The amplitude center was calculated for each 30-second interval within a fraction, and the mean and SD of these centers were subsequently calculated from all fractions.

Systematic and random errors of tumors and OARs

The systematic error (Σ), and random error (σ) of the intrafractional variations of the amplitude center of the tumors and OARs were calculated for all fractions. The Σ was calculated as the SD of the mean amplitude center for each individual patient, and σ as the root mean square of the SD of an individual patient’s amplitude center of the tumor and OARs. To account for geometric uncertainties during treatment delivery, planning margins for both the tumor and OARs were calculated in the LR, AP, SI_coronal, and SI_sagittal directions based on the Σ and σ. The margins for determining the PTV were calculated using van Herk’s formula as follows⁶:

$$2.5\Sigma +0.7\sigma$$ (1)

On the other hand, the margins for determining the PRV were calculated using McKenzie’s formula as follows²⁸:

$$1.3\Sigma +0.5\sigma ,$$ (2)

These formulas account for both the systematic displacement and random blurring effects observed in tumors and OARs during treatment.

Statistical analysis

The statistical significance of differences in ranges of motion between tumors and OARs was evaluated using the Wilcoxon signed-rank test with Holm-Bonferroni correction. All statistical analyses were conducted using R software version 4.4.0, with the threshold for statistical significance set at P = .05.²⁹

Results

Motion range

The median monitoring time was 628.8 seconds (range, 352.2-1087.8 seconds) for all fractions. Figure 2 shows violin plots of the motion ranges of the tumors and OARs. The median motion ranges of the tumor in the LR, AP, SI_coronal, and SI_sagittal directions were 1.8 (IQR, 1.4-2.4), 1.8 (IQR, 1.2-2.5), 5.8 (IQR, 4.5-7.4), and 5.8 (IQR, 4.6-7.2) mm, respectively, which was smaller than those of all OARs (P < .05). There were no statistically significant differences in motion ranges between tumors located in the pancreatic head and those located in the body (P > .05, Figure E2).

Figure 2.

Violin plots of the motion ranges of the tumor and organs at risk in the left-right (LR), anterior-posterior (AP), superior-inferior (SI) in the coronal plane (SI_coronal), and SI in the sagittal plane (SI_sagittal) directions, centered on the tumor (a square with a side length of 10 cm). The number of fractions included in the analysis is indicated in parentheses in the horizontal axis labels. *P <.05.

Abbreviations: fr = fraction.

Among the OARs, colon (SI_coronal: median, 10.2; IQR, 7.8-13.0 mm and SI_sagittal: median, 9.7; IQR, 8.3-11.7 mm) and small bowel (SI_coronal: median, 9.2; IQR, 7.6-11.5 mm and SI_sagittal: median, 8.9; IQR, 7.4-10.8 mm) exhibited larger ranges of motion compared to the duodenum (SI_coronal: median, 7.6; IQR, 5.9-9.6 mm and SI_sagittal: median, 6.3; IQR, 4.7-8.0 mm) and stomach (SI_coronal: median, 7.2; IQR, 6.0-8.7 mm and SI_sagittal: median, 7.1; IQR, 5.8-8.6 mm) (P < .05). For both tumors and OARs, the SI_coronal and SI_sagittal ranges of motion were significantly greater than those in the LR and AP directions (P < .05). A representative case of 2D cine images and motion displacement over the time elapsed of the tumor and OARs in the coronal and sagittal planes is shown in Figure 3. The validation of optical flow-based motion calculation is presented in Figure E3, confirming the appropriateness of this approach for pancreatic MRgART. In this analysis, median 95% Hausdorff distance values for the tumor and OARs ranged from 1.25 to 1.81 mm, while median Dice similarity coefficient values ranged from 0.91 to 0.98 in the coronal and sagittal planes, indicating strong spatial agreement between the manually delineated and optical flow-propagated contours. In addition, when we focused on a narrower region centered on the tumor (a square with a side length of 4 cm), the number of observable OARs decreased; however, for all OARs anatomically adjacent to the tumor, the motion remained comparable to that observed when the broader region (a square with a side length of 10 cm) was used, as shown in Figure E4.

Figure 3.

Results of optical flow calculations in the representative case. (a) Two-dimensional (2D) cine magnetic resonance (MR) images in the coronal and sagittal planes. (b) The time-dependent displacements in the left-right (LR), anterior-posterior (AP), superior-inferior (SI) in the coronal plane (SI_coronal), and SI in the sagittal plane (SI_sagittal) directions. The red arrow within the region of interest of the tumor and organs at risk (OARs) indicates large movements of more than 4 pixels, and the green arrow indicates small movements of less than 4 pixels. The initial contours of the tumor and OARs are displayed in white, and the displaced contours, which were derived from the calculated displacements, are shown in organ-specific colors.

Abbreviation: PTV = planning target volume.

Analysis of baseline drift

Intrafractional variation in the baseline drift is shown in Figure E5 for the LR and AP directions and Figure 4 for the SI_coronal and SI_sagittal directions. Variations of >5 mm in the amplitude center of the tumor at the 30-second, 5-minute, and 10-minute time points, which corresponded to the displacement occurring during the subsequent 30-second intervals, were as follows: 1.2% (4/345 fractions), 11.3% (39/345 fractions), and 22.9% (64/279 fractions) in the SI_coronal direction and 0.9% (3/345 fractions), 12.5% (43/345 fractions), and 26.5% (74/279 fractions) in the SI_sagittal direction, respectively.

Figure 4.

Intrafractional analysis of the baseline drift in (a) the superior-inferior (SI) in the coronal plane (SI_coronal) and (b) SI in the sagittal plane (SI_sagittal) directions for the tumor and organs at risk (OARs). The baseline drifts were calculated as the differences in the amplitude center of the motion range calculated from 2.5% to 97.5% of the displacements at 30-second intervals. Each plot shows the mean of the mid-positions, with error bars representing the SD. The graph titles show the number and proportion of fractions that were included out of the total 345 fractions. The number of fractions at each timepoint is shown in parentheses below the horizontal axis. For both the tumor and OARs, the SD increased with the time that had elapsed, which can be attributed to a gradual reduction in the included fractions.

Abbreviation: fr = fraction.

For OARs, at 10 minutes, which approached the median monitoring time, the proportions of amplitude center variations exceeding ±5 mm in the SI_coronal and SI_sagittal directions were 23.1% (63/273 fractions) and 34.0% (72/212 fractions) for the duodenum, 21.3% (42/197 fractions) and 24.8% (65/262 fractions) for the stomach, 25.3% (64/253 fractions) and 70.0% (21/30 fractions) for the small bowel, and 54.0% (54/100 fractions) and 33.9% (20/59 fractions) for the colon, respectively. The mean amplitude center in the SI_coronal and SI_sagittal directions ranged from −0.92 to −0.23 mm, which indicated a baseline drift toward the inferior direction for both the tumor and OARs.

Correlation between tumors and OARs

The scatterplots of the range of motion between tumors and OARs are illustrated at 30-second intervals with the linear regression line in Figure E6 for the LR and AP directions, and in Figure 5 for the SI_coronal and SI_sagittal directions. The SI_coronal and SI_sagittal motion range between the tumor and OARs were highly correlated (R ≥ 0.89). The slope of the linear approximation of all OARs was >1, which indicated that the ranges of motion of the OARs were greater than those of the tumors. Among the OARs, the colon showed the largest root mean square error values when compared with tumor displacement (Figure E7).

Figure 5.

Plots of the range of motion of the tumor and organs at risk (OARs) in (a) the superior-inferior (SI) in the coronal plane (SI_coronal) and (b) SI in the sagittal plane (SI_sagittal) directions. The slope of the linear approximation was >1, indicating that the motion ranges of OARs were greater than those of the tumor in both SI_coronal and SI_sagittal directions. The graph titles show the number and proportion of included fractions out of the total 345 fractions.

Systematic and random errors of tumors and OARs

Table 1 summarizes the systematic and random errors and the calculated planning margins for the PTV and PRV. Both systematic and random errors were greater in the SI_coronal and SI_sagittal directions than in the LR and AP directions for both tumors and OARs. Consequently, the calculated PTV and PRV margins were largest in the SI_coronal and SI_sagittal directions. The larger margin between the SI_coronal and SI_sagittal directions was as follows: 3.4 mm for the tumor, 2.3 mm for the duodenum, 2.2 mm for the stomach, 2.5 mm for the small bowel, and 2.9 mm for the colon.

Table 1.

Planning margins for the tumors and organs at risk

Variable	Direction	Systematic error (Σ) (mm)	Random error (σ) (mm)	Margin (mm)
Tumor	LR	0.3	0.3	0.9
	AP	0.5	0.4	1.5
	SIcoronal	1.1	0.7	3.1
	SIsagittal	1.1	1.0	3.4
Duodenum†	LR	0.3	0.2	0.5
	AP	0.7	0.3	1.0
	SIcoronal	1.4	0.4	2.1
	SIsagittal	1.5	0.7	2.3
Stomach†	LR	0.4	0.2	0.6
	AP	0.5	0.2	0.7
	SIcoronal	1.5	0.5	2.1
	SIsagittal	1.5	0.6	2.2
Small bowel†	LR	0.5	0.3	0.7
	AP	0.6	0.2	0.8
	SIcoronal	1.6	1.0	2.5
	SIsagittal	1.3	1.1	2.3
Colon†	LR	0.8	0.2	1.2
	AP	0.9	0.2	1.3
	SIcoronal	1.8	1.0	2.8
	SIsagittal	1.8	1.0	2.9

Abbreviations: AP = anterior-posterior; LR = left-right; SI_coronal =superior-inferior in the coronal plane; SI_sagittal = superior-inferior in the sagittal plane.

*Margins for the planning target volume were derived according to van Herk’s formula, defined as 2.5Σ + 0.7σ.⁶

^†Margins for the planning organ at risk volumes were calculated using McKenzie’s formula, defined as 1.3Σ + 0.5σ.²⁸

Discussion

This study investigated the intrafractional anatomic variations in both pancreatic tumors and surrounding OARs using 2D cine MR images acquired during MRgART under AC. By applying an optical flow algorithm to both the tumor and OARs regions, our results confirmed that the OARs exhibited larger motion ranges than the tumors, particularly in the SI direction. Notably, the colon and small bowel exhibited greater motion than did the duodenum and stomach, which are anatomically close to the tumor. These findings support the necessity of individualized, direction-specific margin adaptation strategies combined with careful multiplanar monitoring to achieve high-precision and personalized pancreatic MRgART. Furthermore, our findings provide a comprehensive in vivo perspective on the spatial and temporal relationships among the upper abdominal organs, which can further inform the refinement of motion prediction models and ART strategies.

Previous studies have independently assessed the ranges of motion of pancreatic tumors and adjacent OARs using 2D cine MR images (Table E2).^{8,9,18,19,21,22} However, these studies did not simultaneously examine the interplay between tumors and OARs, and most were limited to a small number of fractions (maximum: 65 fractions). Given the anatomic proximity and potential interactions between tumors and OARs, evaluating concurrent intrafractional variations is critical for minimizing motion-related dosimetric uncertainties. To address this knowledge gap, we analyzed the intrafractional variations and correlation between pancreatic tumors and adjacent OARs, such as the duodenum, stomach, small bowel, and colon, using a large data set of 345 treatment fractions acquired from 25 patients who underwent pancreatic MRgART under AC. Consistent with previous reports,^8,9 our calculated ranges of motion under AC were smaller than those in the free-breathing condition, especially in the SI direction.^18,19

Although previous studies primarily focused on individual OARs such as the small bowel,²² the present study evaluated critical gastrointestinal OARs. Similar to our study, previous studies evaluating the motion of the duodenum, stomach, small bowel, and colon relied on 3D magnetic resonance imaging images that were acquired both pretreatment and posttreatment, and lacked real-time monitoring of OARs during beam delivery.^20,21 Importantly, we found that the ranges of motion of the pancreatic tumors were significantly smaller than those of the adjacent OARs, particularly in the SI direction (Fig. 2). Furthermore, strong correlations between the ranges of motion of tumor and OARs indicate a synchronized respiratory-driven motion pattern during treatment (R ≥ 0.89, Fig. 5). However, the slope of the linear approximation consistently exceeded 1 in the SI direction, which suggests that the OARs tended to exhibit greater motion ranges than the tumor itself. These findings emphasize the importance of an individualized margin-determination strategy that considers the organ-specific motion characteristics of each patient.

The clinical benefit of AC in radiation therapy for pancreatic tumors has been investigated in several studies.30, 31, 32, 33, 34 AC reduces interfractional and intrafractional variations in 2D cine MR^31,32 and CBCT images,³³ and minimizes setup errors.³⁴ The limited intrafractional motion of pancreatic tumors observed in this study can be attributed to the application of AC directly above the tumor site (Fig. 2; median motion range of 5.8 mm in both SI_coronal and SI_sagittal directions). Given their location in the retroperitoneal space, pancreatic tumors are anatomically stabilized by rigid surrounding structures, including the spine and major blood vessels.³⁵ Although AC can reduce the distance between the pancreas and adjacent OARs, the tumor itself was also compressed, and the peripancreatic fat layer was generally thin in this cohort; therefore, the relative spatial relationship between them may not substantially change even without the use of compression.

Furthermore, we found that the small bowel and colon exhibited the greatest motion ranges (median, 8.9-10.2 mm) whereas the duodenum and stomach showed smaller displacements (median, 6.3-7.6 mm) in the SI direction (Fig. 2). These findings align with previous reports which showed that mobile hollow organs, such as the small bowel and colon, exhibit greater positional variability owing to their peristaltic activity and gas content.³⁶ The reduced intrafractional variations in the sagittal plane may be partly explained by the restricted motion of the transverse colon and small bowel located caudal or ventral to the tumor when using AC. For the stomach, the coronal planes captured the pylorus, whereas the sagittal planes visualized the posterior part of the stomach located anterior to the pancreas, suggesting a semifixed position under the AC. The duodenum shows plane-dependent motion patterns, with the coronal plane frequently enabling the visualization of the bulb and descending part, which tends to exhibit larger displacements during respiration. However, the sagittal plane typically includes the horizontal portion that is anatomically fixed behind the pancreas. Given the direction- and plane-dependent motion of OARs adjacent to the tumor, an accurate dose evaluation must account for these variations. This is supported by a previous study that reported that dose deviations do not occur only in the SI direction.³⁶ Therefore, even under AC, differences in OAR motion patterns and visibility between the monitoring planes may affect motion evaluation and adaptive decision-making, highlighting the importance of multiplanar monitoring in pancreatic MRgART.

Baseline drift is another important consideration for intrafractional motion during radiation therapy.⁵ Our results showed that both tumors and OARs exhibited gradual SI shifts with increasing SDs over time (Fig. 4), which is consistent with previous findings obtained using real-time orthogonal kV x-ray imaging in the free-breathing condition.³⁷ Specifically, at 10 minutes, the proportions of amplitude center variations exceeding ±5 mm in the SI direction ranged from 21.3% to 70.0% (Fig. 4). This was partly attributable to the reduced number of fractions in which the OARs remained visible as well as a consistent inferior drift during cine MR acquisition, which is likely driven by progressive muscle relaxation or gas movement in the supine position. Such drift can be problematic in longer treatment sessions, where the cumulative drift may exceed the planned PTV or PRV margins. Grimbergen et al¹¹ suggested the dosimetric benefits of gating and intrafractional drift correction by calculating accumulated planned doses based on comprehensive motion management for upper abdominal tumors. However, their analysis primarily focused on tumor motion without evaluating OAR displacement, as addressed in our study. The clinical implication is that the motion margins should be carefully defined based on real-time monitoring data, as isotropic margins may underestimate geometric uncertainties.

Based on our analysis of systematic and random errors, we derived direction-specific PTV and PRV margins using formulas that were established by van Herk’s⁶ and McKenzie’s formulas,²⁸ respectively (Table 1). Akimoto et al³⁷ investigated intrafractional variations of the 3D pancreas tumor position under free-breathing condition and showed Σ of 1.0, 1.3, and 1.9 mm and σ of 0.5, 0.6, and 0.9 mm in the LR, AP, and SI directions, respectively. In comparison, our results showed Σ of 0.3, 0.5, 1.1, and 1.1 mm and σ of 0.3, 0.4, 0.7, and 1.0 mm in the LR, AP, and SI_coronal, and SI_sagittal directions, respectively (Table 1). Although both 4-dimensional computed tomography and 2D cine MR images are available in our clinical workflow, the lack of sufficient clinical evidence has previously limited our practice to application of uniform isotropic margins from the CTV to the PTV (3-5 mm) on the basis of our institutional protocol.^23,24 The present study, which leveraged a large cohort of patients with accumulated MRgART data, enabled the quantitative evaluation of intrafractional variations in both the tumor and adjacent OARs. Consequently, our evaluation allowed us to estimate population-based margins that more appropriately reflect these motion characteristics and provide a benchmark for future patient-specific margin adaptation.

Phipps et al³⁸ reported that patients are largely able to maintain breath-holding positions within a 3-mm margin, which supports the use of a 3-mm gating margin on the sagittal plane during pancreatic MRgART. Despite the dimensional limitations of 2D cine MR data, our findings suggest that AC effectively suppresses intrafractional variations, allowing for safe margin reduction. Furthermore, the SI margins required for the small bowel and colon were substantially larger than those required for the duodenum or stomach, reflecting higher frame-to-frame variability and more pronounced drift. Rusu et al³⁹ proposed a surrogate gating strategy for the Elekta Unity MR-Linac system in which the GTV was expanded superiorly by 3 cm to extend to the liver edge, creating a high-contrast surrogate structure to compensate for poor tumor visibility on balanced turbo field echo images. However, this approach did not include an assessment of the synchronization between the tumor and surrogate, necessitating margin settings that account for their potential correlation. As demonstrated in our study, this transition toward radiation therapy that incorporates OAR motion using optical flow analysis can enhance the treatment accuracy and safety of pancreatic MRgART.

In addition to their implications for margin determination in MRgART, the motion characteristics identified in this study may offer useful reference information for other radiation therapy modalities. Although the MRgART workflow involves longer treatment times and continuous real-time imaging, whereas CBCT-based treatments have shorter delivery times and lack intrafractional monitoring, the time-resolved motion profiles and baseline drift patterns demonstrated in our analysis may help to inform the expected behavior of pancreatic targets and adjacent OARs under AC. These early-phase motion characteristics are particularly relevant for CBCT-guided ART, given its shorter beam-on times than those of MRgART.

This study has some limitations that should be acknowledged. The first is the use of 2D cine MR data to evaluate the motion of tumors and OARs. Although 2D imaging enables real-time monitoring with a high temporal resolution, it lacks information on out-of-plane motion. However, owing to the technical constraints of the Unity MR-Linac system, which allows real-time monitoring only in fixed planes, our evaluation was limited to the in-plane motion captured by the available 2D cine MR images. Although real-time 3D magnetic resonance imaging reconstruction from 2D cine MR images has been reported for the liver,⁴⁰ further studies are required to investigate the differences in pancreatic tumors and OARs between 2D and 3D MR data. The second limitation is that tumor displacement was assessed without separating the primary tumor from the involved lymph nodes. According to our institutional protocol, the CTV includes both components and is isotropically expanded by 3 to 5 mm in all directions to generate the PTV.^23,24 Because the PTV was created through this uniform expansion, the centroids of the CTV and PTV were essentially identical. Therefore, use of the PTV centroid for motion assessment was considered appropriate for estimating the required margins for both the tumor and adjacent OARs during pancreatic MRgART. Another limitation is that this study used OAR contours that were delineated on the initial frame of 2D cine MR for motion analysis across all frames. This approach does not account for the nonrigid and 3D characteristics of organ motion, which potentially underrepresents anatomic deformation during treatment and limits the accuracy of motion quantification and subsequent dose estimation. Therefore, incorporating deformable contour propagation or volumetric imaging could enhance the fidelity of motion characterization in future studies. Finally, optical flow tends to overestimate the motion range when substantial changes in image intensity occur, especially owing to the intrafractional motion of normal tissues, such as blood vessels. In our analysis, when blood vessels were located adjacent to the tumor, the estimated range of motion occasionally became an outlier. Depending on the tumor location, it may be more reliable to monitor the motion of the surrounding organs that are less affected by intensity variations than the tumor itself.³⁹ Nevertheless, despite these limitations, our optical flow-based approach retains its clinical value for assessing intrafractional motion during pancreatic MRgART.

Conclusions

Optical flow-guided analysis of intrafractional variations revealed that pancreatic tumors exhibited smaller ranges of motion in the SI direction under AC than surrounding OARs, including the small bowel, colon, duodenum, and stomach, despite their proximity. Notably, the colon and small bowel showed especially large and variable SI motions compared with the tumor and other OARs. Our analysis revealed a tendency for baseline drift in the inferior direction over time, with greater variation observed in OARs than in tumors. These results highlight the necessity of accounting for direction-dependent and organ-specific motion characteristics when determining the appropriate PTV and PRV margins for real-time multiplanar monitoring using pancreatic MRgART.

Disclosures

None.