Measuring interfraction and intrafraction lung function changes during radiation therapy using four-dimensional cone beam CT ventilation imaging

Abstract

Purpose:

Adaptive ventilation guided radiation therapy could minimize the irradiation of healthy lung based on repeat lung ventilation imaging (VI) during treatment. However the efficacy of adaptive ventilation guidance requires that interfraction (e.g., week-to-week), ventilation changes are not washed out by intrafraction (e.g., pre- and postfraction) changes, for example, due to patient breathing variability. The authors hypothesize that patients undergoing lung cancer radiation therapy exhibit larger interfraction ventilation changes compared to intrafraction function changes. To test this, the authors perform the first comparison of interfraction and intrafraction lung VI pairs using four-dimensional cone beam CT ventilation imaging (4D-CBCT VI), a novel technique for functional lung imaging.

Methods:

The authors analyzed a total of 215 4D-CBCT scans acquired for 19 locally advanced non-small cell lung cancer (LA-NSCLC) patients over 4–6 weeks of radiation therapy. This set of 215 scans was sorted into 56 interfraction pairs (including first day scans and each of treatment weeks 2, 4, and 6) and 78 intrafraction pairs (including pre/postfraction scans on the same-day), with some scans appearing in both sets. VIs were obtained from the Jacobian determinant of the transform between the 4D-CBCT end-exhale and end-inhale images after deformable image registration. All VIs were deformably registered to their corresponding planning CT and normalized to account for differences in breathing effort, thus facilitating image comparison in terms of (i) voxelwise Spearman correlations, (ii) mean image differences, and (iii) gamma pass rates for all interfraction and intrafraction VI pairs. For the side of the lung ipsilateral to the tumor, we applied two-sided t-tests to determine whether interfraction VI pairs were more different than intrafraction VI pairs.

Results:

The (mean ± standard deviation) Spearman correlation for interfraction VI pairs was ${\overline{r}}_{\mathrm{Inter}}=0.52\pm 0.25$, which was significantly lower than for intrafraction pairs (${\overline{r}}_{\mathrm{Intra}}=0.67\pm 0.20$, p = 0.0002). Conversely, mean absolute ventilation differences were larger for interfraction pairs than for intrafraction pairs, with $\left|\mathrm{\Delta }{\overline{V}}_{\mathrm{Inter}}\right|=0.42\pm 0.65$ and $\left|\mathrm{\Delta }{\overline{V}}_{\mathrm{Intra}}\right|=0.32\pm 0.53$, respectively (p < 10⁻¹⁵). Applying a gamma analysis with ventilation/distance tolerance of 25%/10 mm, we observed mean pass rate of (69% ± 20%) for interfraction VIs, which was significantly lower compared to intrafraction pairs (80% ± 15%, with p ∼ 0.0003). Compared to the first day scans, all patients experienced at least one subsequent change in median ipsilateral ventilation ≥10%. Patients experienced both positive and negative ventilation changes throughout treatment, with the maximum change occurring at different weeks for different patients.

Conclusions:

The authors’ data support the hypothesis that interfraction ventilation changes are larger than intrafraction ventilation changes for LA-NSCLC patients over a course of conventional lung cancer radiation therapy. Longitudinal ventilation changes are observed to be highly patient-dependent, supporting a possible role for adaptive ventilation guidance based on repeat 4D-CBCT VIs. We anticipate that future improvement of 4D-CBCT image reconstruction algorithms will improve the capability of 4D-CBCT VI to resolve interfraction ventilation changes.

1. INTRODUCTION

Currently 10%–30% of lung cancer radiation therapy (RT) patients experience radiation induced lung pneumonitis,^1–3 which represents a major limitation on the prescribed treatment dose. Recently, there has been evidence that lung ventilation imaging (VI) can improve the prediction for radiation pneumonitis,^4,5 which has driven the investigation of ventilation guided RT that minimizes irradiation of healthy lung by incorporating VI into treatment planning.^6–9 It is known that the efficacy of ventilation guidance based on pretreatment VIs is eroded by ventilation changes during treatment, for example, due to tumor regression.^10–12 “Adaptive” ventilation guidance could account for such changes based on repeat VIs during treatment. However, the efficacy of adaptive ventilation guidance is similarly eroded if ventilation differences between short-interval scans, for example, due to patient breathing variability or uncorrelated image noise, are so large that they mask our ability to resolve changes over different weeks of treatment. It is therefore of interest to compare ventilation changes measured over “interfraction” (week-to-week) versus “intrafraction” (pre- and postfraction) timescales.

A challenge is that the gold standard method for ventilation imaging—single photon emission CT (SPECT)¹³—is not convenient for obtaining repeated pre- and postfraction imaging over the course treatment. To address the need for such a technology, we have developed four-dimensional cone beam CT ventilation imaging (4D-CBCT VI), a novel functional lung imaging modality for the radiation therapy treatment room. Anatomic 4D-CBCT provides a convenient basis for adaptive ventilation guidance because it allows the synchronous acquisition of kV projections and respiratory signal on the treatment linear accelerator (“Linac”) just prior to and following treatment delivery.^14–16 4D-CBCT VIs are produced based on existing techniques developed for 4D-CT,^17–19 wherein regional air-volume changes are computed from regional lung deformation using deformable image registration (DIR). The basic steps of adaptive ventilation guided RT based on 4D-CBCT VI are shown in Fig. 1: (i) prefraction 4D-CBCT images are acquired on the treatment Linac, (ii) DIR is used to derive a motion field (arrows) between the exhale and inhale phase images (purple/green overlay), (iii) the regional deformation is then treated as a surrogate for ventilation, thus enabling (iv) adaptive reoptimization of the treatment plan to reduce dose to well-ventilated lung.

FIG. 1.

Potential workflow for adaptive ventilation guided RT using 4D-CBCT VI.

Short-interval reproducibility of 4D-CT VI is reduced by patient breathing variability and we expect the same to be true for 4D-CBCT VI. For example, validation studies of 4D-CT VI have achieved strong voxelwise Spearman correlations (r > 0.8) with contrast-enhanced Xe-CT in mechanically ventilated sheep,^19,20 but for free-breathing humans, the correlations with SPECT and positron emission tomography (PET) ventilation have proved generally weaker (0.2 ≤ r ≤ 0.5).^19,21,22 Du et al. relate this to spontaneous changes in breathing amplitude, frequency, and breathing mode (i.e., abdominal or thoracic) that occur during free-breathing that reduce the reproducibility of regional lung expansion estimates between repeat 4D-CT scans.²³ Patient breathing variability may be mitigated by normalization of repeat 4D-CT VIs based on principles of equivalent tidal volume²⁴ and by maintaining consistent breathing maneuvers,²⁵ for example, using audiovisual breathing guidance.

The reproducibility of 4D-CT VIs has also been linked to underlying 4D-CT image quality. Latifi et al.²⁶ showed that increasing the level of Gaussian image noise in 4D-CT images reduces the concordance of high-function volumes derived using DIR. For 4D-CBCT, an additional consideration is the reconstruction “streak” artifacts arising due to kV projection undersampling.^27,28 These streaks can exhibit 4D motion and similar gray values to lung tissue, thus interfering with the derived motion fields. Streaking artifacts may be countered by applying a number of technologic DIR advancements, including (i) a hierarchical (multiresolution) approach using image downsampling and smoothing to reduce the impact of small spatial structures, (ii) analytic regularization²⁹ to ensure the derived motion fields are spatially smooth, and (iii) self-adaptive cyclic correction³⁰ to filter out streak-induced errors based on the assumption that the respiratory motion is cyclic. To give the reader an appreciation of 4D-CBCT image quality, we point out that the exhale/inhale overlay images of Fig. 1 represent relatively high image quality for 4D-CBCT. It is not unusual for clinical 4D-CBCT scans to suffer from lower tissue contrast, truncated field of view, and more significant streaking artifacts than in this case.

In this work, we use 4D-CBCT to perform the first comparison of interfraction and intrafraction ventilation changes for 19 locally advanced non-small cell lung cancer (LA-NSCLC) patients during RT. This is represented schematically in Fig. 2. For each patient, we generate multiple pre- and postfraction VIs which are deformably registered to a common frame of reference (the time-averaged 4D-CT used for treatment planning), facilitating both segmentation and interfraction/intrafraction comparisons. For the image comparisons, we focus on three metrics. Two of these—the Spearman correlation and mean image difference—describe the overall image similarity (or image difference) as calculated on a voxel level. Meanwhile, a three-dimensional (3D) gamma analysis is used to define the degree of image similarity at different spatial and intensity tolerance levels. Of the three metrics employed, we expect the gamma analysis to be the most useful for expressing the sensitivity of 4D-CBCT VI to interfraction and intrafraction changes.

FIG. 2.

Schematic overview of this study. Process repeated for each patient. Fx = fraction.

Our hypothesis is that patients undergoing lung cancer radiation therapy should exhibit larger interfraction ventilation changes compared to intrafraction function changes. This is based on the assumption that intrafraction VI pairs should represent the case of null-ventilation change. Given a lack of comparable SPECT imaging, we cannot test this assumption directly. Rather intrafraction ventilation changes are minimized as much as possible, by combining a number of advancements including audiovisual breathing guidance, streak-resistant DIR, and global normalization of repeat ventilation scans to account for differences in overall breathing effort. We anticipate that this work represents an important step toward establishing the efficacy of adaptive ventilation guidance and the validity of 4D-CBCT VI.

2. MATERIALS AND METHODS

2.A. Patient data

Nineteen patients with locally advanced NSCLC received a total of 468 4D-CBCT scans on institutional review board-approved protocols at Virginia Commonwealth University (VCU).^31,32 In order to account for different imaging schedules between patients, we limited our study to 215 4D-CBCT scans which for each patient includes 3–4 scans spaced at approximately fortnightly intervals (14 ± 2 days) and between 1–7 scan pairs measured on the same day (34 ± 7 min apart). All patients also received a planning 4D-CT scan. All 4D-CBCT and 4D-CT scans took place between 2008 and 2012 as part of 3D conformal radiotherapy and involved the use of audiovisual biofeedback to coach patient breathing and reduce irregularities in respiration. Table I indicates the primary tumor location and stage for each patient.

TABLE I.

Patient characteristics. RUL =  right upper lobe. RLL  =  right lower lobe. RML  =  right middle lobe. LUL=  left upper lobe. LLL  =  left lower lobe. L  =  liters. N/A  =  not available. ^a To obtain corresponding patient ID in VCU database, add 99 (patients 1–12) or 100 (patients 13–19). ^b Lung volume in time-averaged 4D-CT (planning CT). ^c Fraction of lung common to all 4D-CBCT VIs after alignment to planning CT.

Patient IDa	Stage	Tumor location	Segmented volume (L)b	Fraction imaged (%)c
1	IIIB	RUL	4.6	73
2	IIIB	RUL	4.0	64
3	IIIB	LLL	4.0	81
4	IIIB	RUL	3.7	67
5	IIIA	LLL	3.2	93
6	IIIA	LLL	4.6	86
7	IIIA	Mediastinum	2.0	72
8	IIIA	LUL	2.1	86
9	IIIA	RUL	5.0	59
10	IIIA	RLL	3.2	84
11	IIIA	RML	4.6	85
12	IIIA	RML	3.4	88
13	IIIA	RUL	3.6	61
14	IIIA	RML/LL	6.8	75
15	IIIA	RUL	3.0	71
16	IIIA	LUL	4.8	61
17	IIIA	RML	2.1	99
18	IIIA	RML	1.4	93
19	IIIA	RUL/ML	4.0	57

4D-CBCT scans were acquired using an On-Board Imager (Varian Medical Systems) with in-house hardware and software modifications. For each scan, approximately 2400 projections were acquired over a period 8–10 min (about 100 breathing cycles) in a single 360° slow gantry arc. Projections were acquired in half-fan mode using a half bowtie filter with x-ray tube settings of 125 kVp, 20 mA, and 20 ms. Projections were then sorted into six phase bins by extracting a respiratory signal based on the sum of attenuation values in the lower-left quadrant of each projection image, roughly corresponding to the diaphragm location. More details on this technique can be found in Ref. 28. Each phase-correlated projection set was reconstructed using the well-known Feldkamp–Davis–Kress (FDK) backprojection algorithm³³ as implemented in COBRA (Cobra Exxim, Inc.). All reconstructed 4D-CBCT phase images had a slice thickness of 2 mm, with 448 × 448 axial resolution and 0.88 × 0.88 mm² pixel size.

4D-CTs were acquired using a 16-slice helical CT scanner (Brilliance Big Bore; Philips Medical Systems) with a slice thickness of 3 mm and 512 × 512 axial resolution (0.98 ×  0.98 mm² pixel size). Contours for the primary tumor were generated on the time-averaged 4D-CT or “planning CT” in Pinnacle (Philips Medical Systems). Lungs were segmented on the planning CT using an in-house threshold-based approach with intensity cutoff −500 Hounsfield units (HUs).

2.B. Generation of 4D-CBCT VIs

4D-CBCT VIs were produced using our in-house ventilation image software VESPIR (“VEntilation, Segmentation and Pulmonary Image Registration”). VESPIR provides a matlab interface to the B-spline DIR software Plastimatch (http://plastimatch.org), which we have previously validated for 4D-CT.²¹ In this study, we have applied an additional temporal filtering of the motion fields to reduce the impact of 4D-CBCT streak artifacts. To achieve this, we implemented a “self-adaptive cyclic correction” DIR scheme similar to that developed by Brehm et al.,³⁰ which was originally applied to motion-compensated enhancement of 4D-CBCT phase images. For detailed descriptions of the DIR parameters, and our implementation of the streak-reduced DIR approach, we point the reader toward Appendices A and B, respectively.

Briefly, as an initial step DIR was performed between each pair of adjacent 4D-CBCT phase images I_ϕ (“fixed” image) and I_ϕ+1 (“moving” image) for phase number ϕ = 1,  …,  N. Similar to our earlier 4D-CT study (Ref. 21), and based on the findings of Appendix A, we used a dual-resolution B-spline control grid as shown in Table II. Here, the DIR cost function employed a mean square error (MSE) image similarity metric and a motion field regularization λ value of 5. The output of each DIR process is a motion field u_(ϕ,ϕ+1) that maps points from I_ϕ into I_ϕ+1. After N registrations, we obtain a “daisy-chain” of motion fields,

$${\displaystyle \{{\mathbf{u}}_{(1,2)},{\mathbf{u}}_{(2,3)},...,{\mathbf{u}}_{(N-1,N)},{\mathbf{u}}_{(N,1)}\}.}$$ (1)

It is at this stage that we apply the self-adaptive cyclic correction of Appendix B, the impact of which is evidenced in Fig. 3(a). In the presence of 4D-CBCT streaking artifacts, self-adapting cyclic correction DIR can visibly improve the reproducibility of motion fields between high- and low-motion cases for the same patient.

TABLE II.

Parameters for two-stage B-spline DIR used in this study. Both stages used regularization λ = 5.

Stage	B-spline grid (mm3)	Image downsampling	No. of iterations
1 (“Coarse”)	18 × 18 × 20	4 × 4 × 2	1000
2 (“Fine”)	9 × 9 × 10	2 × 2 × 1	500

FIG. 3.

(a) Effect of streak-resistant DIR on 4D-CBCT motion fields; cyclic correction reduces spurious motion outside the patient. (b) Effect of DIR λ on heterogeneity of derived regional ventilation. See detailed discussion in Appendix B.

Based on the corrected motion fields, we compose a motion field between the exhale and inhale phases, v, which is the motion field associated with ventilation. The ventilation V is then quantified as the fractional exhale-to-inhale lung volume change at voxel x, written V(x) = (J(x) − 1), where J(x) is the Jacobian determinant of v,

$${\displaystyle \mathrm{J}\left(\mathbf{x}\right)=\left|\begin{array}{rrr}\hfill 1+\frac{\partial {\mathbf{v}}_x\left(\mathbf{x}\right)}{\partial x}& \hfill \frac{\partial {\mathbf{v}}_x\left(\mathbf{x}\right)}{\partial y}& \hfill \frac{\partial {\mathbf{v}}_x\left(\mathbf{x}\right)}{\partial z}\\ \hfill \frac{\partial {\mathbf{v}}_y\left(\mathbf{x}\right)}{\partial x}& \hfill 1+\frac{\partial {\mathbf{v}}_y\left(\mathbf{x}\right)}{\partial y}& \hfill \frac{\partial {\mathbf{v}}_y\left(\mathbf{x}\right)}{\partial z}\\ \hfill \frac{\partial {\mathbf{v}}_z\left(\mathbf{x}\right)}{\partial x}& \hfill \frac{\partial {\mathbf{v}}_z\left(\mathbf{x}\right)}{\partial y}& \hfill 1+\frac{\partial {\mathbf{v}}_z\left(\mathbf{x}\right)}{\partial z}\end{array}\right|.}$$ (2)

It is worth considering the sensitivity of the Jacobian metric to interfraction ventilation changes measured using DIR. Ideally, the smallest detectable change relates to the maximum deformation when performing DIR between identical fixed/moving images. In our case, when performing DIR between identical 4D-CBCT phase images, the 99th percentile of Jacobian values (including image boundary regions) was 1.02, equivalent to a 2% regional volume difference. In practice however, the lower limit of sensitivity will be set by the patient-specific intrafraction breathing variability and it is difficult to make a prediction in this regard.

For lung cancer patients, one surrogate of breathing variability is the root mean square error (RSME) of tumor displacement from a mean breathing cycle. For a mean peak-to-peak displacement of 5 mm, a RMSE of 1 mm or 20% during an imaging or treatment session is not uncommon and could imply similar levels of intrafraction variations in regional ventilation. We note also that the Plastimatch B-spline control points can be spaced no closer than 4 voxels or 10 mm given our image downsampling schedule in Table II. So as a first order estimate, we can expect that the smallest detectable ventilation change should be in the range 2%–20% over distances ≥10 mm.

We point out that alternate ventilation metrics, for example, based on regional HU changes as in Refs. 18, 21, and 22, were not analyzed due to the non-HU equivalence of the 4D-CBCT images.

2.C. Alignment of fortnightly VIs

In order to compare ventilation in the presence of changing patient anatomy and changing scanner field of view, all 4D-CBCT VIs were deformably registered to the corresponding planning CT, segmented based on the planning CT lung structures and cropped to a common region of interest. This was achieved by converting each anatomic 4D-CBCT image to a streak-reduced, 3D CBCT representing the time-average of the 4D-CBCT phase images. Each CBCT underwent a rigid + deformable registration to the corresponding planning CT. The DIR approach was similar to that of Sec. 2.B except for the use of a mutual information (MI) image similarity metric and the exclusion of tumor voxels from the cost function (to reduce registration errors arising from nonconservation of tumor mass between fractions). After the alignment/segmentation procedure, all VIs underwent a 3 mm median filter to eliminate any small nonconnected lung structures. The rightmost columns of Table I compare the volumes of lung imaged in the planning CTs and aligned VIs.

2.D. Normalization of 4D-CBCT VIs

Each VI was normalized based on the functional percentile values for relatively “healthy” lung, i.e., the lung contralateral (opposite) to the tumor. This was achieved by applying a linear scaling

$${\displaystyle {}^{\ast }V\left(\mathbf{x}\right)=(V\left(\mathbf{x}\right)-V_{10}^{\mathrm{contra}})/(V_{90}^{\mathrm{contra}}-V_{10}^{\mathrm{contra}}),}$$ (3)

where ^∗V is the normalized ventilation at voxel x, with $V_{10}^{\mathrm{contra}}$ and $V_{90}^{\mathrm{contra}}$ the 10th and 90th functional percentile values for contralateral lung, respectively. Equation (3) has the effect of rescaling each image between the 10th/90th functional percentiles, which take values of 0 and 1, respectively. This normalization approach was chosen to emphasize the highest/lowest ventilation regions in each image. In cases where the 10th/90th functional percentile values are similar in absolute terms, this may exaggerate regional ventilation differences compared to the raw ventilation map. As such, our measurements of ventilation change may similarly represent an upper limit of the true physiological ventilation change. All tumor voxels and other nonlung voxels were assigned with a value −1000 and excluded from analysis.

2.E. Interfraction and intrafraction image comparisons

For each patient, we collated the intrafraction VI pairs (including pre/postfraction scans on the same day) and interfraction VI pairs (including first day scans and each of treatment weeks 2, 4, and 6) using the notation “Intra,” “Week2,” “Week4,” and “Week6,” respectively. Interfraction pairs contained only prefraction scans, with some of these also used in intrafraction pairs. Since the interfraction and intrafraction ventilation changes are potentially confounded by interfraction DIR errors, we also compared interfraction and intrafraction CBCT image pairs after registration to the planning CTs. To this end, CBCTs were normalized similarly to Eq. (3), using the 10th and 90th intensity percentile values of each CBCT image (corresponding to the approximate intensities of air and water). The variability of ipsilateral lung voxels in both VIs and CBCTs was computed using the three following metrics, each incorporating a total of ∼5 × 10⁷ ipsilateral lung–voxel pairs across the 19 patients.

2.E.1. Spearman correlations

Spearman correlation values r^{V I} (or r^CBCT) were obtained for each set of VI (CBCT) pairs. These take a value between −1 and 1 and represent the degree of monotonicity of intensities in spatially correlated voxels. The mean correlations were compared using two-sided t-tests, with a significance level 0.05 and a null hypothesis that the mean r^{V I} (or r^CBCT) for intrafraction and interfraction pairs was the same.

2.E.2. Normalized ventilation differences

We produced difference maps between each VI (and CBCT) pair, obtaining intrafraction and interfraction ventilation differences ΔV (intensity differences ΔI). These values were compared using two-sided t-tests, again with a significance level 0.05 and a null hypothesis that there was no difference in $\mathrm{\Delta }\overline{V}$ (or $\mathrm{\Delta }\overline{I}$) between the intrafraction and interfraction image pairs.

2.E.3. Gamma pass rates

We used Plastimatch to generate 3D gamma distributions for all intrafraction VI and CBCT pairs. The gamma distribution provides a difference map incorporating both normalized intensity differences and physical distance to agreement (“DTA”) on a combined unitless scale.³⁴ We tested the similarity of intrafraction VI (and CBCT) pairs using a range of different tolerance criteria for intensity (3%, 5%, 10%, 25%, 50%) and DTA (3 mm, 5 mm, and 10 mm). Ipsilateral lung voxels with a gamma value ≥1 indicate a “fail” of one or both similarity criteria. For intrafraction VI (CBCT) pairs, we identified the smallest tolerance criteria providing a median gamma pass rate of 80%, deemed to be an arbitrarily “acceptable” level of agreement. This was compared to the gamma pass rates for interfraction VI (CBCT) pairs based on that same criteria.

2.F. Summary of image comparison methodology

Before presenting our results, let us briefly summarize our methodology above. Our goal is to compare and contrast patient-specific changes in functional VIs and anatomic CBCT images, for ipsilateral lung, over both interfraction and intrafraction timescales. To facilitate the interfraction comparisons, the VIs and anatomic CBCTs for each patient are mapped to the reference frame of the corresponding planning CT and then scaled to the intensity range 0–1 based on Eq. (3). For VIs, this accounts for differences in overall breathing effort; for anatomic CBCTs, this accounts for scan-to-scan variations in the intensity levels for air/water. In addition to calculating the median ventilation for ipsilateral lung, the similarity of interfraction and intrafraction image pairs is quantified using three metrics: (i) the voxel-based Spearman correlations, (ii) normalized image differences, and (iii) gamma pass rates based on different tolerance levels for intensity differences and DTA.

Our hypothesis is that patients undergoing lung cancer radiation therapy should exhibit larger interfraction ventilation changes compared to intrafraction ventilation changes. In order to support our hypothesis, it is necessary to demonstrate that interfraction VI pairs are more different than intrafraction VI pairs. Additionally, in order to discount differences arising from interfraction DIR errors, it is also necessary to show that the change in image similarity between interfraction and intrafraction VI pairs is larger than the change in image similarity between interfraction and intrafraction anatomic CBCT image pairs.

3. RESULTS

3.A. General observations of interfraction ventilation change

Figure 4 shows all the ipsilateral VIs included in the interfraction comparisons, meanwhile Fig. 5 shows the first intrafraction VI pair for each patient. All VIs are plotted on a intensity scale of 0 − 1 relative to the contralateral lung function [cf. Eq. (3)]. Subjectively, we can identify a broad range of longitudinal behaviors. Three patients showed a clear ventilation increase and three showed a decrease. Seven patients exhibit “transient” changes in ventilation magnitude, while still maintaining a consistent spatial distribution. Patient 3 exhibits low (but stable) ventilation, whereas patient 2 exhibits “erratic” changes in the distribution of high/low ventilated lung. Interestingly, in comparing Figs. 4 and 5, two of the patients (11 and 17) show erratic interfraction changes but still exhibit at least one example of good intrafraction reproducibility.

FIG. 4.

Complete set of ipsilateral VIs used for interfraction comparisons. VIs (inside lung region) are superimposed onto CBCT images after deformable registration to planning CT and grouped into subjective categories. All images are shown in the axial midplane of the tumor (solid contour). *Ventilation relative to contralateral lung.

FIG. 5.

Collection of VIs representing the first intrafraction VI pair for each patient; images have been subjectively grouped in terms of visual reproducibility. *Ventilation relative to contralateral lung.

Table III shows the fortnightly values of (i) the median contralateral ventilation, $V_{50}^{\mathrm{contra}}$, prior to any global normalization, and (ii) median ipsilateral ventilation, $V_{50}^{\mathrm{ipsi}}$, after global normalization by contralateral lung [cf. Eq. (3)]. It is interesting to compare the interfraction variability of unnormalized $V_{50}^{\mathrm{contra}}$ and normalized $V_{50}^{\mathrm{ipsi}}$, as these values should capture changes in breathing effort and underlying function, respectively. From the table, we see that both $V_{50}^{\mathrm{contra}}$ and $V_{50}^{\mathrm{ipsi}}$ exhibit large interfraction changes. For seven patients, the mean interfraction change in $V_{50}^{\mathrm{contra}}$ exceeded 25% (bolded entries in the table), underscoring the need to account for breathing effort differences visible in healthy lung.

TABLE III.

Fortnightly values of (i) $V_{50}^{\mathrm{contra}}$ prior to any global normalization and (ii) $V_{50}^{\mathrm{ipsi}}$ normalized by contralateral lung [cf. Eq. (3)]. %ΔV refers to the percent ventilation change between first day scans and each of weeks 2, 4, and 6. STD is standard deviation. Bolded entries indicate mean changes greater than 25%.

	(i) $V_{50}^{\mathrm{contra}}$ (without normalization)					(ii) $V_{50}^{\mathrm{ipsi}}$ [normalized by Vcontra; see Eq. (3)]
Patient ID	First scan	Week 2	Week 4	Week 6	Mean% ΔV	First scan	Week2	Week4	Week6	Mean% ΔV	Max% ΔV	(Week #)
1	0.09	0.13	0.06	0.08	+1.6	0.19	0.32	0.76	0.45	+170	+300	(4)
2	0.08	0.08	0.08	0.08	−3.6	0.18	0.29	0.24	0.24	+42	+61	(2)
3	0.09	0.09	0.06	0.04	−29	−0.25	−0.44	−0.08	−0.05	−23	+80	(6)
4	0.14	0.13	0.12	0.09	−22.0	0.57	0.43	0.47	0.36	−26	−36	(6)
5	0.11	0.11	0.13	0.13	+11	0.36	0.11	0.33	0.56	−5.9	−69	(2)
6	0.06	0.13	0.11	0.06	+59	0.33	0.27	0.16	0.25	−32	−52	(4)
7	0.09	0.19	0.17	0.15	+82	0.58	0.30	0.46	0.43	−31	−47	(2)
8	0.11	0.07	0.09	0.06	−32	0.29	0.67	0.29	0.66	+89	+130	(2)
9	0.07	0.05	0.06	0.05	−21	0.31	0.67	0.17	0.43	+34	+110	(2)
10	0.02	0.09	0.06	0.05	+230	1.29	0.22	0.44	0.59	−68	−83	(2)
11	0.11	0.12	0.07	0.15	+1.5	0.64	0.63	0.87	0.44	+1.5	+37	(4)
12	0.11	0.08	0.09	0.10	−21	0.41	0.15	0.13	−0.04	−81	−110	(6)
13	0.07	0.05	0.07	0.05	−18	0.53	0.43	0.47	0.53	−10	−19	(2)
14	0.04	0.06	0.02	0.03	−6.8	0.57	0.48	0.68	0.73	+11	+28	(6)
15	0.14	0.07	0.08	0.10	−37	0.22	0.46	0.43	0.31	+82	+110	(2)
16	0.06	0.06	0.07	0.06	−0.10	0.29	0.34	0.17	0.25	−14	−41	(4)
17	0.12	0.09	0.13	0.09	−12	0.26	0.14	0.22	0.27	−19	−46	(2)
18	0.21	0.17	0.17	0.20	−17	0.00	0.02	0.12	0.16	+3800	+6200	(6)
19	0.10	0.07	0.07	—	−25	0.43	0.27	0.58	—	−2.4	−39	(2)
Mean	0.10	0.10	0.09	0.08	+7.16	0.38	0.30	0.36	0.37	+210
± STD	0.04	0.04	0.04	0.05	60.6	0.31	0.26	0.24	0.21	880

After normalizing the ventilation images according to Eq. (3), we observed that the median ipsilateral ventilation sits between the 10th and 90th functional percentiles of contralateral lung (i.e., $0\le V_{50}^{\mathrm{ipsi}}\le 1$) for most scans; averagedover all interfraction scans, we found ${\overline{V}}_{50}^{\mathrm{ipsi}}=0.35\pm 0.26$. Ten patients exhibited mean absolute changes in $V_{50}^{\mathrm{ipsi}}\ge 25\%$ (bolded entries in the table); of these patients, five showed a ventilation decrease and five showed an increase. For all patients, the mean change had the same direction as the largest change, with the maximum change occurring at weeks 2, 4, and 6 around 50%, 25%, and 25% of the time, respectively.

3.B. Interfraction and intrafraction comparisons

3.B.1. Spearman correlations

Figure 6(a) shows a boxplot of Spearman correlations for interfraction and intrafraction pairs of ipsilateral VIs (shaded boxes) and anatomic CBCTs (white boxes). Qualitatively, interfraction VI pairs show lower correlations than intrafraction VI pairs; VI pairs also show lower correlations than corresponding CBCT image pairs. However, the difference in correlations between interfraction and intrafraction image pairs is similar for both VIs and CBCTs. In other words, while interfraction VI changes do exceed intrafraction VI changes, we cannot rule out that some component of interfraction VI variability is due to changes in anatomic position, which are not completely resolved by DIR of the fortnightly CBCTs to the planning CT.

FIG. 6.

Boxplots of (a) Spearman correlations and (b) absolute normalized intensity differences for different VI pairs (shaded boxes) and CBCT image pairs (white boxes). For each box, the median, lower, and upper 25th percentile values are shown. Whiskers denote the range of all nonoutlier values as determined by matlab. Data incorporate all ipsilateral lung voxels across 19 patients.

Quantitatively, the (mean ± standard deviation) correlation for intrafraction VI pairs was ${\overline{r}}_{\mathrm{Intra}}^{\mathrm{V\; I}}=0.67\pm 0.20$. The correlations between first day VIs and each of weeks 2, 4, and 6 were markedly lower (${\overline{r}}_{\mathrm{Week2}}^{\mathrm{V\; I}}=0.51\pm 0.26$, ${\overline{r}}_{\mathrm{Week4}}^{\mathrm{V\; I}}=0.54\pm 0.22$, and ${\overline{r}}_{\mathrm{Week6}}^{\mathrm{V\; I}}=0.51\pm 0.27$, respectively). Merging all the interfraction data together, we obtained ${\overline{r}}_{\mathrm{Inter}}^{\mathrm{V\; I}}=0.52\pm 0.25$; this was significantly different to ${\overline{r}}_{\mathrm{Intra}}^{\mathrm{V\; I}}$ (p = 2 × 10⁻⁴).

The variability of CBCT pairs was generally lower than for the corresponding VI pairs; we obtained ${\overline{r}}_{\mathrm{Intra}}^{\mathrm{CBCT}}=0.87\pm 0.05$, ${\overline{r}}_{\mathrm{Week2}}^{\mathrm{CBCT}}=0.79\pm 0.12$, ${\overline{r}}_{\mathrm{Week4}}^{\mathrm{CBCT}}=0.78\pm 0.10$, and ${\overline{r}}_{\mathrm{Week6}}^{\mathrm{CBCT}}=0.78\pm 0.10$. Merging all the interfraction data, we obtained ${\overline{r}}_{\mathrm{Inter}}^{\mathrm{CBCT}}=0.79\pm 0.10$ which was significantly different to ${\overline{r}}_{\mathrm{Intra}}^{\mathrm{CBCT}}$ (p ∼ 10⁻⁹). As noted above, the difference in correlations between interfraction and intrafraction image sets was not too different between VIs and CBCTs, implicating a potential role of interfraction DIR errors in accounting for interfraction ventilation changes; we measured $({\overline{r}}_{\mathrm{Intra}}^{\mathrm{V\; I}}-{\overline{r}}_{\mathrm{Inter}}^{\mathrm{V\; I}})=0.15$ and $({\overline{r}}_{\mathrm{Intra}}^{\mathrm{CBCT}}-{\overline{r}}_{\mathrm{Inter}}^{\mathrm{CBCT}})=0.09$.

3.B.2. Normalized image differences

Figure 6(b) shows a boxplot of absolute ventilation differences |ΔV| (shaded boxes) and CBCT intensity differences |ΔI| (white boxes) for ipsilateral lung. VI pairs show larger interfraction differences than intrafraction differences; we found $\left|\mathrm{\Delta }{\overline{V}}_{\mathrm{Intra}}\right|=(0.32\pm 0.53)$, $\left|\mathrm{\Delta }{\overline{V}}_{\mathrm{Week2}}\right|=0.42\pm 0.67$, $\left|\mathrm{\Delta }{\overline{V}}_{\mathrm{Week4}}\right|=0.43\pm 0.66$, and $\left|\mathrm{\Delta }{\overline{V}}_{\mathrm{Week6}}\right|=0.40\pm 0.62$. The merged interfraction differences were $\left|\mathrm{\Delta }{\overline{V}}_{\mathrm{Inter}}\right|=(0.42\pm 0.65)$, which were significantly different to the intrafraction differences (p < 10⁻¹⁵).

As with the VI pairs, CBCT pairs showed larger interfraction differences compared to intrafraction differences; we found $\left|\mathrm{\Delta }{\overline{I}}_{\mathrm{Intra}}\right|=(0.05\pm 0.05)$, $\left|\mathrm{\Delta }{\overline{I}}_{\mathrm{Week2}}\right|=0.09\pm 0.10$, $\left|\mathrm{\Delta }{\overline{I}}_{\mathrm{Week4}}\right|=0.08\pm 0.08$, and $\left|\mathrm{\Delta }{\overline{I}}_{\mathrm{Week6}}\right|=0.08\pm 0.08$. The merged interfraction differences were $\left|\mathrm{\Delta }{\overline{I}}_{\mathrm{Inter}}\right|=(0.08\pm 0.09)$, significantly different to $\left|\mathrm{\Delta }{\overline{I}}_{\mathrm{Intra}}\right|$ (p < 10⁻¹⁵). Compared to the Spearman correlations of Sec. 3.B.1, interfraction changes in mean absolute ventilation were less confounded by interfraction DIR error; we measured $\left(\right|\mathrm{\Delta }{\overline{V}}_{\mathrm{Inter}}|-|\mathrm{\Delta }{\overline{V}}_{\mathrm{Intra}}\left|\right)=0.10$ compared to $\left(\right|\mathrm{\Delta }{\overline{I}}_{\mathrm{Inter}}|-|\mathrm{\Delta }{\overline{I}}_{\mathrm{Intra}}\left|\right)=0.03$.

3.B.3. Gamma pass rates

Figure 7(a) shows gamma pass rates for intrafraction VI pairs (shaded symbols) and CBCT pairs (white symbols) for intensity tolerance criteria in the range 3%–50%. The different shapes indicate different DTA criteria (3, 5, or 10 mm). For intrafraction CBCTs, we observed mean gamma pass rates ≥80% for all tested combinations of intensity/distance tolerance. For VIs however, only a subset of these combinations provided pass rates above 80%, namely, the combinations 25%/10 mm, 50%/3 mm, 50%/5 mm, and 50%/10 mm which yielded gamma pass rates 80% ± 15%, 87% ± 12%, 88% ± 12%, and 92% ± 9%, respectively.

FIG. 7.

(a) Gamma pass rates for intrafraction VI pairs (shaded symbols) and CBCT pairs (white symbols) for a range of tolerance criteria for intensity differences (3%–50%) and DTA (3–10 mm). (b) Mean gamma pass rates for interfraction VI pairs (black symbols) and intrafraction VI pairs (gray symbols) for each patient using intensity/distance tolerance of 25%/10 mm. Error bars (where shown) indicate standard deviation.

In order to test interfraction variations, we chose intensity/distance tolerance criteria of 25%/10 mm, as this represents smallest intensity tolerance producing mean gamma pass rates ≥80% between intrafraction VI pairs. Based on this, interfraction VI pairs had gamma pass rates 77% ± 16% compared to ∼99% for CBCTs. Compared to intrafraction image pairs, interfraction gamma pass rates were significantly different for VIs (p = 0.0003) but not for CBCTs (p ∼ 0.5).

Figure 7(b) compares the mean interfraction and intrafraction gamma pass rates separately for each patient. Mean interfraction gamma pass rates were lower than intrafraction pass rates for 15 out of 19 patients, indicating that VI reproducibility was better over short intervals than between weeks. Ten of these patients correspond to the bolded entries of Table III showing the largest changes in the median ipsilateral ventilation between fractions. Conversely, of the five patients showing mean interfraction gamma pass rates larger than intrafraction pass rates, none corresponded to large interfraction changes in Table III. In other words, we observed that interfraction ventilation changes exceeded intrafraction changes where large interfraction changes were involved.

4. DISCUSSION

Our data support the hypothesis that lung cancer patients can exhibit greater interfraction ventilation changes compared to intrafraction ventilation changes on average. Importantly, these interfraction changes were measured after applying a global normalization to account for patient breathing variability. Based on Table III, we have observed that lung cancer patients can exhibit both positive and negative ventilation changes throughout conventional RT treatment, moreover this maximum change may occur in different weeks for different patients. These results are significant because they indicate the potential for adaptive ventilation guidance to minimize healthy lung irradiation compared to both conventional RT or ventilation guided RT based on a single pretreatment scan.

At first glance, the lack of a clear common pattern for interfraction ventilation changes in Fig. 4 is surprising. It seems that more or less for every patient one could make the point that “ventilation” is sometimes more, sometimes less, varying during treatment. Interestingly, this observation is in line with Vinogradskiy et al.,¹² who measured functional dose–response of different lung lobes based on weekly 4D-CT VIs but did not find a consistent pattern of ventilation change. Those authors did find that ventilation tends to increase for lobes containing shrinking tumors. While we have not stratified our results in terms of the tumor volume change, we can make a similar observation that the median (normalized) ventilation for ipsilateral lung appears lower than for contralateral lung.

Also surprising is that, in Fig. 5, we see a few examples where intrafraction VI pairs seem more different than interfraction VI pairs. This could be explained if the patient spontaneously changes from thoracic to abdominal breathing between pre- and postfraction scans, but it could also be that in some cases, the breathing motion is so small that the DIR remains driven by image noise/artifacts. In any case, it is important to quantify the sensitivity of 4D-CBCT VI to genuine interfraction changes, after the effects of breathing variability are masked out (in our case, by applying a global normalization based on the ventilation of contralateral lung). Based on our 3D gamma analyses of the normalized ipsilateral ventilation, we suggest that 4D-CBCT VI can detect interfraction changes exceeding a tolerance of 25%/10 mm. For these tolerance criteria, it is possible to observe a significant difference in gamma pass rates between interfraction VI pairs (cases of anticipated ventilation change) compared to intrafraction VI pairs (anticipated null-change). Generally, interfraction ventilation changes exceeded intrafraction changes for cases where large changes in the median ipsilateral ventilation (e.g., ≥25%) were also observed to occur. In other words, adaptive ventilation guidance based on repeat 4D-CBCT VIs may be possible for patients experiencing severe ventilation changes during treatment. Determining what constitutes a “severe” ventilation change will require correlation of 4D-CBCT VI changes against spirometry-based pulmonary function tests and toxicity outcomes. This will be the subject of future work.

A limitation of our analysis is that the interfraction DIR between CBCTs and planning CTs—used to align interfraction VIs—suffers limited accuracy in the parenchyma near the tumor. This is because the DIR focuses on high-contrast boundaries as opposed to the parenchyma structure which has a lower contrast, particularly in CBCT. We attempt to limit this issue by masking out nonlung voxels in both the DIR and ventilation analysis, but even so the effect of mismatching near the tumor may propagate outward due to a combination of DIR regularization and lack of contrast in adjacent mismatched regions. Based on the variability of deformed CBCT images, we estimate that up to 50% of the measured interfraction VI variability may be attributable to interfraction DIR errors. However, some of that variability is likely also due to streaking artifacts and tissue perfusion changes.

Another limitation is that our implementation of self-adaptive DIR [cf. Eqs. (B1) and (B2) in Appendix B] assumes an equal magnitude of registration error between any two phases, which is not necessarily the case. The implementation by Brehm et al.³⁰ proposed higher order corrections in the form of additional iterations of Eqs. (B1) and (B2). For example, the cyclic correction could first be performed between all adjacent phase bins, then every second phase bin, every third phase bin, and so on. Irrespective of such higher order corrections, Fig. 3(a) shows that even a “first order” correction, as in this work, may improve the reproducibility of motion fields between high- and low-motion 4D-CBCT scans for the same patient.

Given that this is the first extensive work on 4D-CBCT VI, it is also interesting to compare the quality of 4D-CBCT ventilation as compared to 4D-CT. As shown in Fig. 8, 4D-CBCT VIs look similar to 4D-CT VIs generated using the same DIR parameters and the same Jacobian-based ventilation metric. In a few cases, we can spot disparities in the distribution of ventilation between upper/lower or left/right lung between 4D-CBCT VIs and 4D-CT VIs. While it might seem natural to assume that 4D-CT always provides the “correct” image, it is important to remember that both 4D-CT and 4D-CBCT ventilation can suffer image artifacts related to irregular motion. For cine-mode 4D-CT, anatomic truncation/duplication at abutting couch positions could cause erroneous bands of ventilation bright/dark spots in the superior–inferior (SI) direction. For 4D-CBCT however, our primary concern is typically the motion of reconstruction streak artifacts perpendicular to the SI direction. A question for both 4D-CBCT and 4D-CT ventilation is whether Jacobian-based ventilation is meaningful for all patients; the expansion/contraction of lungs may not always correlate to gas exchange going on in those areas (for example, in emphysematous bullae). Indeed, there is evidence that in free-breathing humans, intensity-based ventilation metrics correlate better with nuclear medicine than Jacobian-based metrics.²¹

FIG. 8.

Comparison of Jacobian-based 4D-CT VIs and the first 4D-CBCT VI for each patient. VIs (inside lung region) are unnormalized and superimposed onto corresponding CT/CBCT. All images are shown in the coronal midplane of the tumor (solid contour).

We anticipate that there is much potential to improve the quality of 4D-CBCT VI. In the near future, iterative 4D-CBCT reconstruction algorithms could provide a pathway to enable intensity based ventilation metrics and improve the accuracy of DIR, by reducing 4D-CBCT streak artifacts and image noise. In terms of the acquisition itself, our results reinforce the notion that breathing effort differences should always be considered for different scan times, even between intrafraction scans. The reproducibility of intrafraction 4D-CBCT VI may benefit from some form of breathing training or verbal instruction that patients should try to remember the same breathing pattern within and between scans.

5. CONCLUSIONS

We have performed the first comparison of interfraction and intrafraction ventilation changes for lung cancer radiation therapy patients using 4D-CBCT VI. We found that interfraction ventilation changes exceed intrafraction changes for cases of large interfraction ventilation change, moreover these patients can exhibit ventilation increase or decrease over the course of treatment. These observations represent an important step toward establishing the efficacy of adaptive ventilation guidance and the validity of 4D-CBCT VI. Additional optimization of 4D-CBCT image reconstruction and DIR algorithms may help to improve the short-interval reproducibility of 4D-CBCT VIs and sensitivity to lung function changes over the course of treatment.

APPENDIX A: OPTIMIZATION OF DIR PARAMETERS

The DIR procedure of Sec. 2.B requires some tuning. Each motion field u of Eq. (1) is chosen to minimize the cost function C(u) = C_MSE(u) + λC_reg(u) representing the elastic nature of lung tissue deformation. Here, C_MSE is the MSE intensity difference between the fixed and deformed moving images, and C_reg is a regularization term relating to the second derivative of u. The scalar λ controls the trade-off between image similarity and motion field smoothness and represents the main parameter to be optimized. A value of λ that is too small results in motion fields driven by image noise, whereas a value of λ that is too large results in over-smooth motion fields that underestimate tissue motion.

For the case where DIR is performed on the entire thoracic image, we can define the optimal λ to be the smallest value for which ventilation signal is confined to the lung anatomy. This is a reasonable criteria given the assumption of incompressible tissue outside the lung. Figure 3(b) demonstrates the impact of λ on the ventilation heterogeneity; based on a visual inspection of first day VIs for all patients, we chose λ = 5. For the ventilation computation in Sec. 2.B, the DIR cost function includes only those voxels within the lungs of the fixed and/or moving images (based on a coarse, threshold-based lung segmentation of both images). This does not appreciably impact the choice of λ.

Similar to our earlier study using 4D-CT,²¹ a Gaussian filter of kernel width 0.5 mm was applied to all 4D-CBCT phase images prior to registration; Gaussian filtering is commonly applied in lung ventilation studies to help reduce the effects of quantum noise. In this case, the 0.5 mm filter has only a small effect given the 0.88 × 0.88 mm² pixel size and does not reduce the problem of streaking artifacts.

APPENDIX B: IMPLEMENTATION OF STREAK-REDUCED DIR METHOD

Here, we describe our implementation of self-adaptive cyclic correction DIR, similar to that developed by Brehm et al.³⁰ As explained in Sec. 2.B, the first step of our DIR process is to obtain a daisy-chain of motion fields, u_(ϕ,ϕ+1), between each pair of adjacent 4D-CBCT phase images, where the phase number is given by ϕ = 1,  ...,  N.

Following this initial step, the motion fields of Eq. (1) are temporally filtered based on the assumption that the time-integrated displacement of any lung tissue element over the set of N motion fields should precisely be zero. That is, we assume that the respiratory motion is exactly periodic. Practically, each motion field u_(ϕ,ϕ+1) is actually associated with some small registration error ϵ_(ϕ,ϕ+1) and so the time-integrated displacement is nonzero. We estimate the overall registration error by composing a motion field starting and ending at the same phase,

$${\displaystyle {\mathbf{U}}_{(\varphi ,\varphi )}={\mathbf{u}}_{(\varphi ,\varphi +1)}\circ {\mathbf{u}}_{(\varphi +1,\varphi +2)}\circ ...\circ {\mathbf{u}}_{(\varphi -2,\varphi -1)}\circ {\mathbf{u}}_{(\varphi -1,\varphi )}\sim N{ϵ}_{(\varphi ,\varphi +1)}.}$$ (B1)

By assuming each ϵ_(ϕ,ϕ+1) is small, we obtain the corrected motion fields

$${\displaystyle {\mathbf{u}}_{(\varphi ,\varphi +1)}^{\mathrm{corrected}}={\mathbf{u}}_{(\varphi ,\varphi +1)}-{ϵ}_{(\varphi ,\varphi +1)}\sim {\mathbf{u}}_{(\varphi ,\varphi +1)}-({\mathbf{U}}_{(\varphi ,\varphi )}/N),}$$ (B2)

with Eqs. (B1) and (B2) applied once for each value of ϕ. We then compose the corrected motion fields between the exhale and inhale phases (e and i, respectively). This is the motion field associated with ventilation, written $\mathbf{v}={\mathbf{U}}_{(e,i)}^{\mathrm{corrected}}$.