Skip to main content
NCBI home page
Wiley Open Access Collection logoLink to Wiley Open Access Collection
. 2025 Jul 15;52(7):e17942. doi: 10.1002/mp.17942

Evaluation of deformable image registration accuracy for liver re‐irradiation patients using contrast and non‐contrast computed tomography images

Caitlin Allen 1,2,3,✉, Adam Yeo 1, Rick Franich 2, Sarat Chander 4,5, Nicholas Hardcastle 1,5,6
PMCID: PMC12260772  PMID: 40660871

Abstract

Background

Re‐irradiation of liver tumors is increasing in frequency, requiring clinicians to account for previous radiation dose to prevent unacceptable toxicity. Given the heterogeneity of liver morphological changes between treatments, deformable image registration (DIR) is required to accumulate dose from previous treatments onto the latest planning images for radiotherapy.

Purpose

The increase in re‐irradiation of patients with liver cancer has led to the need to account for previous radiotherapy treatments. This feasibility study used contemporaneous intravenous contrast computed tomography images (CTs) to evaluate the accuracy of DIR dose accumulation in the re‐irradiation of liver patients, via the use of structural landmarks.

Methods

We used nine liver patients who received repeat stereotactic body radiation therapy (SBRT) liver radiotherapy, with contrast and non‐contrast planning CTs, to evaluate the accuracy of dose accumulation in the liver. The initial planning CT was deformed to the second planning CT, and the deformation vector field was used to deform the initial dose map. The dose could then be accumulated by adding the deformed initial dose map to the second dose map. Three methods of performing DIR were compared, including with and without corresponding anatomical landmarks. Target registration error (TRE), dice similarity coefficient, and Hausdorff distance were used to assess the accuracy of the dose accumulation.

Results

The lowest TRE was achieved with the structure guided algorithm using all of the available anatomical landmarks, with a mean + standard deviation of 1.7 mm (SD = 0.9 mm) for non‐contrast (p < 0.0005) and 1.6 mm (SD = 0.9 mm) for contrast CTs (p < 0.0005). DIR based on contrast CTs reduced the TREs, the distance between the location of a given landmark on the second image, and the location of where that landmark is deformed to from the first image, with all DIR algorithms (p < 0.0005 for each contrast non‐contrast pair). There were also statistically significant differences between dose accumulation errors for Contrast CTs with a mean of 0.92 Gy (SD = 3.08 Gy) and Non‐Contrast CTs of 1.07 Gy (SD = 3.36 Gy) (p < 0.05), and the differences between each of the algorithms were also statistically significant, with p‐values < 0.05.

Conclusions

DIR improves the dose accumulation accuracy in re‐irradiation in liver SBRT. DIR accuracy can be improved using contrast CTs and corresponding anatomical landmarks. Providing additional information into the DIR in the form of corresponding anatomical landmarks dramatically improved image registration accuracy and thus reduced dose accumulation errors. Dose accumulation accuracy was dependent on the TRE, and on the dose‐gradient of the mapped dose.

Keywords: dose accumulation, deformable image registration, liver SBRT re‐irradiation

1. INTRODUCTION

Stereotactic body radiation therapy (SBRT) is an emerging treatment for liver tumors from primary liver cancer and metastases from other primary tumors. SBRT is a hypofractionation therapy that involves delivering a much higher dose to the tumor per fraction. SBRT has been proven to result in both high local control rates and low toxicity. 1 , 2 , 3 , 4 , 5 , 6 , 7 , 8 , 9 , 10 Re‐irradiation is increasing in frequency due to improved long‐term survival of patients with various cancers, increased focus on localized treatment to sites of metastatic disease, and improved ability to minimize dose to critical, previously irradiated organs through use of intensity modulation and image guidance. 11

Although liver SBRT has a high rate of local control (3‐year actuarial local control rates for primary liver tumors of 86% and 76% for liver metastases), 10 there are still up to 24% of tumors that will recur that need further treatment, as well as new disease foci in a previously treated liver. This leads to the need for repeated SBRT treatments, which in turn increases the risk of toxicity. 12 , 13 , 14 , 15 , 16 It is desirable to account for the radiation dose previously delivered to the liver from previous radiotherapy when designing a new treatment plan. Given the large and heterogeneous impact of SBRT on liver size and shape, deformable image registration (DIR) is warranted to improve the accuracy of mapping previously delivered doses between courses. 12 , 17 , 18 , 19 , 20 , 21 , 22 , 23 Image intensity‐based DIR requires a high spatial registration accuracy to ensure voxels are mapped to the correct location to avoid accumulating dose in the wrong region of the liver. The same limitations that make it difficult to achieve a high spatial accuracy for the registration also make it difficult to assess the DIR accuracy; the liver has a homogeneous appearance in the non‐contrast CTs used for treatment planning, making the various liver regions indistinguishable. However, in order to delineate the target, liver patients typically undergo another form of imaging such as a contrast CT. Contrast CT imaging using an iodine contrast agent to image arterial and or portal venous phases is a popular modality often performed at the time of the planning CT, thereby reducing the anatomical differences between the contrast and planning CTs. 24 , 25 , 26 Contrast CT images allow visualization of blood vessels, providing a means of identifying various tissue regions within the liver, which can be used to aid in the quality assurance of and to guide DIR.

The aim of the current work was to conduct a feasibility study to evaluate the accuracy of deformable mapping of SBRT dose from multiple SBRT courses. We compared the accuracy of deformable dose mapping based on non‐contrast and contrast CTs, using anatomical features derived from contemporaneous intravenous contrast CTs. It was hypothesized that including anatomical landmarks in the DIR process would increase the accuracy of the dose accumulation in the liver.

2. METHOD

2.1. Patient selection

Ethics approval was granted by the Peter MacCallum Cancer Centre Ethics Committee. All nine patients who received repeat courses of SBRT for tumors in the liver between April 2018 and March 2021 at a single institution were included in this study, to give a total of thirteen cases. For each plan, each patient underwent a non‐contrast treatment planning CT in exhale breath hold followed immediately by arterial and portal venous phase contrast CTs, also in breath hold, with an iodine intravenous contrast agent. Two patients underwent three courses of SBRT, and their plans were split into three pairs: Courses 1 and 2, Courses 2 and 3, and Courses 1 and 3. Individual patient details are provided in Table 1, which includes diagnosis, the number of targets, its location and volume, normal liver volumes, prescription doses, and time relapse between each course.

TABLE 1.

Individual liver patient details which includes diagnosis, the number of targets, its location and volume, normal liver volumes, prescription doses and time relapse between each course.

Patient (no. landmarks) Diagnosis Course Liver Vol (cm3) Liver Vol minus GTVs (cm3) No. GTVs (segment) Prescribed dose
Liver 1 (24) HCC Course 1 1042 1019 1 (Seg 8) 40 Gy in 5 fx
Course 2 (relapse: 11 mths) 1041 1039 1 (Seg 8) 40 Gy in 5 fx
Liver 2 (17) HCC Course 1 1987 1958 2 (Seg 6 & Seg 7) 48 Gy in 3 fx
Course 2 (relapse: 48 mths) 1867 1837 2 (Seg 5 & Seg 8) 45 Gy in 5 fx
Liver 3 (14) PC met Course 1 1601 1512 3 (Seg 8 & two in Seg 6) 42 Gy in 3 fx
Course 2 (relapse: 5 mths) 1568 1545 2 (Seg 5/8 & Seg 6) 30 Gy in 3 fx
Liver 4 (19) CRC met Course 1 1322 1302 1 (Seg 4A/8) 50 Gy in 5 fx
Course 2 (relapse: 10 mths) 1279 1273 1 (Seg 6) 40 Gy in 5 fx
Liver 5 (19) HCC Course 1 1216 1189 2 (Seg 6 & Seg 7) 45 Gy in 5 fx
Course 2 (relapse: 20 mths) 775 773 2 (Seg 3 & Seg 8) 30 Gy in 5 fx
Liver 6 (21) CRC met Course 1 1475 1463 1 (Seg 6) 48 Gy in 3 fx
Course 2 (relapse: 3 mths) 1549 1545 2 (Seg7 & Seg 8) 48 Gy in 3 fx
Liver 7 (22) CRC met Course 1 1070 1070 1 (Seg 5) 48 Gy in 3 fx
Course 2 (relapse: 4 mths) 1072 1035 2 (Seg 3 & Seg 8) 45 Gy in 3 fx
Liver 8 (22–29) HCC Course 1 694 690 1 (Seg 5) 30 Gy in 5 fx
Course 2 (relapse: 1 mths) 925 921 1 (Seg 8) 40 Gy in 5 fx
Course 3 (relapse: 6 mths) 747 742 1 (Seg 5) 40 Gy in 5 fx
Liver 9 (19–23) MTC met Course 1 893 878 2 (Seg 7 & Seg 8) 42 Gy in 3 fx
Course 2 (relapse: 3 mths) 857 754 1 (Seg 7/8) 42 Gy in 3 fx
Course 3 (relapse: 6 mths) 979 961 1 (Seg 8) 42 Gy in 3 fx
Open in a new tab

Abbreviations: CRC, colorectal cancer; HCC, hepatocellular cancer; MTC, medulary thyroid cancer; Met, metastasis; PC, pancreatic cancer; Seg, segment.

2.2. Landmark selection

Corresponding landmarks were identified on each pair of portal venous contrast images based on blood vessel bifurcations and the presence of any surgical clips or calcifications. Landmark identification was performed by one observer, with review and refinement by two independent observers, to reduce interobserver variability. All three observers were medical physicists. A rigid registration (RIR) was performed between the contrast and non‐contrast CTs in Eclipse (v15.6, Varian Medical Systems, Palo Alto, CA), and the landmark points were transferred to the non‐contrast image. The RIR was verified by visual inspection of the liver outline and any landmarks visible on both images such as clips or calcifications. Vessel landmark locations were assumed to be mapped accurately due to the lack of visibility of these in the non‐contrast images.

2.3. DIR algorithms used

DIR was performed using the VelocityAI software (v4.1, Varian Medical Systems, Palo Alto, CA). 27 VelocityAI features a multiresolution b‐spline algorithm combined with the mutual information similarity index. 27 We investigated two deformable registration algorithms: the deformable multi‐pass (DMP), and the structure‐guided (SG) algorithms. DMP performs deformable registration at increasingly finer image resolution for each pass, to a total of three passes, and is suitable for CT to CT registration. 27 The SG algorithm is a hybrid technique that applies a higher weighting of the cost function in the region surrounding structures contoured on both input images in order to guide the DMP algorithm. 27 The SG algorithm is recommended for more extreme anatomical changes when there is substantial change in the volume of corresponding structures. 27

2.4. DIR procedure

Before performing DIR, a RIR was first performed between each pair of non‐contrast planning CTs, and each pair of contrast planning CTs, in Eclipse. To do this, a rectangular region of interest (ROI) was used that corresponded to approximately 1 cm expansion of the liver contour from the first plan. The goal was to ensure that the liver borders from both plans were included in the RIR, while minimizing the amount of surrounding anatomy included. This RIR was then used as the starting point for all subsequent DIR.

An ROI was used for the DIRs by setting the ROI to maximum patient extent in every direction except in the superior, inferior, and left‐hand side of the patient, which was set to 2 mm greater than the greatest extent of the liver from either CT. The SG algorithm was performed in one of two ways: using a subset of five landmarks selected throughout the liver to guide the registration or using all of the available landmarks for a given patient. In every case, the Course 1 CT (moving) was registered to the Course 2 CT (fixed). For the purposes of clinical implementation, five landmarks were suggested as a realistic compromise between time required to annotate if performed in a clinical workflow, and registration accuracy. This process was performed for both non‐contrast pairs, and portal‐venous contrast pairs to determine the impact of improved internal liver anatomy visualization on DIR accuracy. A flow chart of this process is shown in Figure 1.

FIGURE 1.

FIGURE 1

Open in a new tab

A flow chart of the basic liver dose accumulation method. The red dots are a visual representation of the landmarks used in the study.

2.5. Evaluation of DIR accuracy

To assess the DIR accuracy, the target registration error (TRE), the Dice Similarity Coefficient (DSC), and the Hausdorff distance (HD) were calculated. The TRE is the distance between the location of a given landmark on the second image and the location of where that landmark is deformed to from the first image. The contour‐based metrics were calculated using the liver contours obtained from the treatment plans. The DSC is a measure of the degree of overlap of the liver contours as per the second planning CT and deformed from the first planning CT, whereby 0 refers to no overlap, and 1 is perfect overlap. The HD is a measure of the maximum distance to agreement between the contour surfaces. 28 , 29 , 30 , 31

2.6. Dose accumulation

Dose accumulation was performed in Velocity by applying the deformation vector fields (DVFs) obtained from the DIR to the dose map from the first plan to warp it to the frame of reference of the planning image for the second plan. The deformed dose map could then be added to the second dose map to give the accumulated dose. The calculated dose at each pair of corresponding landmarks in each plan was summed, and the dose accumulation error at each point was calculated as the difference between this and the registration‐based accumulated dose at each landmark. 32 A positive dose accumulation error means the registration‐based accumulated dose is higher than the summation of both points.

2.7. Statistical analysis

The effect sizes of differences in TREs and absolute dose between registration methods were calculated based on the bootstrapped mean difference statistic, using a 95% confidence interval computed using estimation statistics. 33 Wilcoxon Sign Ranked tests for paired data were used to determine if there were statistically significant differences between the DIR of contrast and non‐contrast images, or any of the DIR algorithms used. The Wilcoxon Sign Rank test compares the distribution of values between paired datasets. 34 The null hypothesis was that, for each comparison, there would be no difference in the two datasets, and therefore, there was no benefit to using one registration algorithm over another. The Spearman correlation coefficient was computed to determine any correlation between the TREs, accumulated dose differences, and the dose gradients. 35 In all cases, statistical significance was based on the 95% confidence interval. To reduce the effect of Type I errors resulting from multi‐comparisons, the Holm–Benferroni correction was applied to the statistical analysis.

3. RESULTS

The number of landmarks identified was dependent on the visibility of any features and varied between patients from 14 to 29; see an example of the spread of landmarks for Liver 9 in Figure 2. The TREs for the various registration algorithms are given in Figure 3, along with their bootstrapped mean difference statistics. The RIR resulted in the largest TREs due to an inability to account for soft tissue deformation, with a median [IQR] TRE of 7.1 mm [4.7–11.2 mm] for non‐contrast and 6.6 mm [4.2–9.2 mm] for contrast CTs. DIR with DMP reduced the median TRE to 4.8 mm [3.1–7.1 mm] for non‐contrast and 3.7 mm [2.3–5.6 mm] for contrast CTs (p < 0.0005 for all compared with RIR). Similarly, the SG algorithm using five landmarks further reduced the median TRE to 2.9 mm [1.7–4.7 mm] for non‐contrast and 0.9 mm [1.7–4.0 mm] for contrast CTs (p < 0.0005 for both). The lowest TRE was achieved with the SG algorithm using all of the available landmarks, 0.9 mm [1.2–2.3 mm] for non‐contrast and 0.9 mm [1.2–2.3 mm] for contrast CTs (p < 0.0005 for both). DIR based on contrast CTs reduced the TREs compared to DIR based on non‐contrast CTs with all DIR algorithms (p < 0.0005 for each contrast non‐contrast pair).

FIGURE 2.

FIGURE 2

Open in a new tab

A 3D visualization to illustrate the spread of landmarks (crosses) throughout the liver volume (yellow) for Liver 9, Course 2. The PTV for Course 2 is shown in blue. Landmarks from both Courses 1 and 2 are shown, with color coding to identify matching pairs. The underlying CT image (Course 2 pCT) shows the coronal mid‐plane of the PTV with the orthogonal mid‐planes pinned. Some landmarks located behind the mid‐plane of the PTV are not shown.

FIGURE 3.

FIGURE 3

Open in a new tab

TRE analysis for registration from (a) contrast CT to contrast CT and (b) non‐contrast to non‐contrast CT. Statistically significant effect size comparisons (p < 0.05) are marked with *.

Absolute differences between the summed point doses and the RIR/DIR accumulated point doses for the various registration algorithms are given in Figure 4, along with their bootstrapped mean difference statistics. While these results follow the same overall pattern as the TRE values, the bootstrapped mean differences are smaller.

FIGURE 4.

FIGURE 4

Open in a new tab

Absolute dose difference analysis for registration from (a) contrast CT to contrast CT and (b) non‐contrast to non‐contrast CT. Statistically significant comparisons (p < 0.05) are marked with *.

The liver contour DSC and HD for the various registration algorithms are given in the Supporting Information. The mean DSCs were between 0.90 and 0.91. The median [IQR] DSC was higher for non‐contrast CTs when registration was based on the DMP algorithm, 0.9 [0.9–1.0], compared with either the SG with five landmarks, 0.9 [0.9–0.9], or the SG with all available landmarks, 0.9 [0.9–0.9], p < 0.05. There were no differences between DSC for the other images and registrations. The median [IQR] HD was between 19.5 [11.7–24.8] for SG with all available landmarks for contrast CTs and 21.2 [11.7–24.8] for the RIR of contrast CTs, with no differences between registration algorithms or contrast / non‐contrast CTs.

Differences between the summed point doses and the RIR/DIR accumulated point doses are depicted as a function of the TRE and the dose gradient in Figures 5 and 6. In all algorithms, the dose accumulation error was lower when using contrast CTs, compared with non‐contrast CTs, p < 0.05. In general, the algorithms appear to follow the same pattern as for the TREs. Based on contrast CTs, the dose accumulation error with RIR was 0.4 Gy [0.0–3.3 Gy], with DMP was 0.1 Gy [0.0–1.5 Gy], with SG (5 landmarks) was 0.1 Gy [0.0–0.7 Gy] and with SG (all landmarks) was 0.1 Gy [0.0–0.3 Gy] (all pairwise comparisons p < 0.05). When using non‐contrast CTs, the dose accumulation error with RIR was 0.4 Gy [0.0–3.6 Gy], with DMP was 0.2 Gy [0.0–2.0 Gy], with SG (5 landmarks) was 0.1 Gy [0.0–0.8 Gy] and with SG (all landmarks) was 0.1 Gy [0.0–0.4 Gy] (all pairwise comparisons p < 0.05). The SG with all available landmarks achieved a maximum absolute dose difference of 6.4 Gy. The dose accumulation error at each landmark as a function of TRE at that landmark is plotted in Figure 5. A weak positive correlation was found between the absolute dose accumulation error and the TRE (r = 0.27, p < 0.005). Furthermore, there appears to be a bias towards the registration‐based dose accumulation being higher than the point dose summation from individual plans, which is more evident for the rigid algorithm.

FIGURE 5.

FIGURE 5

Open in a new tab

Dose accumulation error between landmark point doses accumulated via RIR/DIR and manually summed doses vs. target registration error (TRE) at each point for registration of contrast CT to contrast CT, and non‐contrast CT to non‐contrast CT. A positive dose accumulation error corresponds to the registration‐based accumulated dose being higher than the summation of both points.

FIGURE 6.

FIGURE 6

Open in a new tab

Absolute dose difference between landmark point doses accumulated via RIR/DIR and the dose gradient at each point for registration of contrast CT to contrast CT, and non‐contrast CT to non‐contrast CT. A positive dose accumulation error means the registration‐based accumulated dose is higher than the summation of both points.

Plots of the absolute dose accumulation errors at each landmark as a function of the dose gradient at that landmark are in Figure 6. There was a weak positive correlation between the absolute value of the dose accumulation error and dose gradient (r = 0.33, p < 0.005). A comparison of liver dose metrics (mean dose, V15Gy, and V<15 Gy) for each of the dose accumulations is given in the Supporting Information. In general, the value of the dose metrics varied between the different DIR algorithms, and the SG algorithm using all landmarks typically had the lowest dose value.

4. DISCUSSION

It is important to accurately quantify previously delivered dose to critical organs at the time of treatment planning in a re‐irradiation scenario, and to estimate total organ cumulative doses from repeat courses of radiation therapy. The liver is particularly challenging due to the lack of soft tissue features visible within the liver on non‐contrast CTs, and the large deformations observed in response to treatment. We have shown DIR reduces the TRE and, subsequently, dose accumulation error when mapping dose between CT scans of livers in repeat courses of SBRT, compared with RIR. If DIR is performed using contrast CTs, where internal liver anatomy is better visualized, DIR accuracy is improved, however this did not translate to improved dose accumulation accuracy.

Registration accuracy can be further improved if anatomical landmarks are used to focus the cost function in specific regions. In the structure guided algorithm utilized in this study, the anatomical landmarks do not act as boundary conditions to the DIR but provide regions in which the weight on the cost function is increased in the DIR optimization problem. Therefore, although we do not expect perfect registration accuracy at these landmarks, the registration is expected to improve in regions around these landmarks. As demonstrated by the reduction in the TREs, as shown in Figure 3, our results provide validation that the inclusion of anatomical landmarks in the DIR improves registration accuracy at these locations. We observed a mean difference between the TREs for the DMP and SG algorithm using five landmarks of −1.5 mm for the contrast CTs and −2.5 mm for the non‐contrast CTs, with a further reduction between the TREs for the DMP and SG algorithm using all landmarks of −3.1 mm for the contrast CTs and −4.4 mm for the non‐contrast CTs. This is a significant reduction in the positional error for dose accumulation, considering the typical resolution of the dose calculation grid for SBRT is 1 mm × 1 mm × 1 mm. The ability to assess registration accuracy in regions away from the corresponding landmarks, however, remains a challenge. We have demonstrated improvement of DIR accuracy when using anatomical landmarks; it may not be prudent to exclude any available corresponding landmarks from the registration purely for the purposes of assessing registration accuracy.

We observed a trend of positive dose accumulation errors. It was expected that points located in higher dose gradient regions would also have a higher probability of larger dose differences since SBRT treatments have a steep dose fall‐off in the region surrounding the tumor, but this does not seem to be the case. This could be explained by the TRE having a greater effect or possibly the dose gradient at that location. Alternatively, it could be argued that this effect would have been observable for points immediately surrounding the tumor, but due to difficulties identifying landmarks in a region of the liver that has undergone substantial changes due to tumor regression and radiation damage, this could not be tested. Regardless, the mean difference between the absolute dose differences for the DMP and SG algorithm using five landmarks was −0.4 Gy for the contrast CTs and −0.6 Gy for the non‐contrast CTs, with a further reduction between the TREs for the DMP and SG algorithm using all landmarks of −1.0 Gy for the contrast CTs and −1.1 Gy for the non‐contrast CTs. This demonstrates that the use of anatomical landmarks also improves the accuracy of the dose accumulation.

A comparison table of the TRE results of liver registrations in the literature with the results of our study is given in the Supporting Information. In general, there is a pattern of decreasing TREs with increasing information used to drive the registration. Reported TREs for RIRs have mean values ranging from 7.9 mm to 11.8 mm. 36 , 37 Although DIR usually improves TREs, this is not always the case and reported values range from 4.6 to 12.4 mm. 36 , 37 , 38 , 39 Various techniques can be used to incorporate anatomical information to drive the registration, including landmarks (fiducials, surgical clips, vessel bifurcations), contours, surface boundary conditions, and biomechanical models. Studies using these techniques were able to reduce their mean TREs, ranging from 4.4 to 7.7 mm. 39 , 40 , 41 , 42 Furthermore, in studies that compared the results of registration with non‐contrast and contrast CTs, in all cases, there was a decrease in the mean TRE for contrast CTs. This is the same pattern seen in our results, with the highest mean TRE of 8.4 mm (SD = 4.8 mm) for RIR of non‐contrast CTs, and the lowest mean of 1.6 mm (SD = 0.9 mm) for contrast CTs registered with the SG algorithm using all landmarks.

Although the use of landmarks improved registration accuracy, identification of landmarks is time‐consuming and difficult due to the extent of soft tissue deformation that can occur between the two time points and constitutes an important limitation in terms of clinical implementation. We estimate the uncertainty in point selection to be approximately equal to the slice thickness in the superior‐inferior direction (2 mm) and in the axial direction (1 mm). To introduce this work into the clinic, more efficient procedures for landmark detection are required. One possibility would be to incorporate an automatic means of generating landmarks such as the SIFT algorithm, 43 which could be used to guide the registration, with clinicians identifying a set of landmarks for use in quality assurance. For example, Cazoulat et al. (2020, 2021) used an automated segmentation of liver vasculature to generate sets of corresponding landmarks for TRE evaluation. 37 , 41 Furthermore, even if the points are selected accurately in the contrast CTs, there is still the possibility of introducing errors when mapping points from the contrast CTs to the non‐contrast CTs. We have assumed that the contrast and non‐contrast CTs have the same anatomy, but it is possible for there to be deformation of the liver occurring between the two time points that may not be identifiable through visual inspection. Therefore, when designing a new DIR algorithm, we can conclude that it would be advantageous to incorporate not only a means of including known features to guide the registration, but also an automated means of identifying features in both images prior to the registration.

Another issue is determining how best to perform quality assurance for dose accumulation. It is not covered by the guidelines provided by TG 132 due to the complexities involved. 31 Unlike, other DIR uses, dose accumulation requires that each individual voxel is mapped to the correct location on the primary image as the deformation itself is the desired output. In comparison, when generating a synthetic CT for dose calculations, it does not matter where each individual voxel ended up so long as the final HU is approximately correct. Likewise, for contour propagation, it is the boundary of the contour that is important, and any deformations occurring within an organ will not affect the final outcome. Furthermore, we do not have a known full 3D “ground truth” dose with which to compare our accumulated dose. As such, we decided to base our quality assurance on the point doses since we were able to calculate a gold standard based on the sum of the point doses for the same landmark for each SBRT plan. From this, we can compare the summed point doses with the point doses obtained from the accumulated dose map. The limitation in this case is that we do not know the accuracy of the dose accumulation in the regions surrounding the landmarks. Further uncertainty with respect to the radiobiological impact of deformed dose may also exist due to liver atrophy and hypertrophy, which are scenarios in which current DIR algorithms have shortcomings. 44 , 45 Regardless, we believe the inclusion of anatomical landmarks in the SG algorithm in VelocityAI can improve dose mapping accuracy from a previous SBRT treatment to a new planning CT, to help aid in planning a second SBRT treatment for liver cancer. While the sample size is a limiting factor in our analysis, the patient cohort covers a wide range of displacements and anatomical variations such as liver size, which still allows for the generalization of our results. Future work would include a full clinical study including an analysis of local control rates and incidence of radiation‐induced liver disease.

5. CONCLUSION

DIR improves the dose accumulation accuracy in re‐irradiation in liver SBRT. DIR accuracy can be improved using contrast CTs to delineate corresponding anatomical landmarks. Increasing additional information into the DIR in the form of corresponding anatomical landmarks drastically improved image registration accuracy and reduced dose accumulation error. Dose accumulation accuracy was dependent on the TRE of the registration, and on the dose gradient of the mapped dose.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest.

Supporting information

Supplementary Materials

MP-52-0-s001.pdf (776.7KB, pdf)

ACKNOWLEDGMENTS

The authors acknowledge the support received for this research through the provision of an Australian Government Research Training Program Scholarship. This work was also funded in part by the Peter MacCallum Cancer Foundation.

Open access publishing facilitated by RMIT University, as part of the Wiley ‐ RMIT University agreement via the Council of Australian University Librarians.

Allen C, Yeo A, Franich R, Chander S, Hardcastle N. Evaluation of deformable image registration accuracy for liver re‐irradiation patients using contrast and non‐contrast computed tomography images. Med Phys. 2025;52:e17942. 10.1002/mp.17942

REFERENCES

  • 1. Eriguchi T, Takeda A, Sanuki N, et al. Acceptable toxicity after stereotactic body radiation therapy for liver tumors adjacent to the central biliary system. Int J Radiat Oncol. 2013;85(4):1006‐1011. doi: 10.1016/j.ijrobp.2012.09.012 [DOI] [PubMed] [Google Scholar]
  • 2. Cárdenes HR, Price TR, Perkins SM, et al. Phase I feasibility trial of stereotactic body radiation therapy for primary hepatocellular carcinoma. Clin Transl Oncol Off Publ Fed Span Oncol Soc Natl Cancer Inst Mex. 2010;12(3):218‐225. doi: 10.1007/s12094-010-0492-x [DOI] [PubMed] [Google Scholar]
  • 3. Andolino DL, Johnson CS, Maluccio M, et al. Stereotactic body radiotherapy for primary hepatocellular carcinoma. Int J Radiat Oncol Biol Phys. 2011;81(4):e447‐453. doi: 10.1016/j.ijrobp.2011.04.011 [DOI] [PubMed] [Google Scholar]
  • 4. Price TR, Perkins SM, Sandrasegaran K, et al. Evaluation of response after stereotactic body radiotherapy for hepatocellular carcinoma. Cancer. 2012;118(12):3191‐3198. doi: 10.1002/cncr.26404 [DOI] [PubMed] [Google Scholar]
  • 5. Bujold A, Massey CA, Kim JJ, et al. Sequential phase I and II trials of stereotactic body radiotherapy for locally advanced hepatocellular carcinoma. J Clin Oncol Off J Am Soc Clin Oncol. 2013;31(13):1631‐1639. doi: 10.1200/JCO.2012.44.1659 [DOI] [PubMed] [Google Scholar]
  • 6. Yoon SM, Lim Y‐S, Park MJ, et al. Stereotactic body radiation therapy as an alternative treatment for small hepatocellular carcinoma. PLoS One. 2013;8(11):e79854. doi: 10.1371/journal.pone.0079854 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7. Sanuki N, Takeda A, Oku Y, et al. Stereotactic body radiotherapy for small hepatocellular carcinoma: a retrospective outcome analysis in 185 patients. Acta Oncol Stockh Swed. 2014;53(3):399‐404. doi: 10.3109/0284186X.2013.820342 [DOI] [PubMed] [Google Scholar]
  • 8. Weiner AA, Olsen J, Ma D, et al. Stereotactic body radiotherapy for primary hepatic malignancies—report of a phase I/II institutional study. Radiother Oncol J Eur Soc Ther Radiol Oncol. 2016;121(1):79‐85. doi: 10.1016/j.radonc.2016.07.020 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9. Bujold A, Dawson LA. Stereotactic radiation therapy and selective internal radiation therapy for hepatocellular carcinoma. Cancer/Radiothérapie. 2011;15(1):54‐63. doi: 10.1016/j.canrad.2010.11.003 [DOI] [PubMed] [Google Scholar]
  • 10. Ohri N, Tomé WA, Méndez Romero A, et al. Local control after stereotactic body radiation therapy for liver tumors. Int J Radiat Oncol. 2021;110(1):188‐195. doi: 10.1016/j.ijrobp.2017.12.288 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11. Andratschke N, Willmann J, Appelt AL, et al. European Society for Radiotherapy and Oncology and European Organisation for Research and Treatment of Cancer consensus on re‐irradiation: definition, reporting, and clinical decision making. Lancet Oncol. 2022;23(10):e469‐e478. doi: 10.1016/S1470-2045(22)00447-8 [DOI] [PubMed] [Google Scholar]
  • 12. Lee S, Kim H, Ji Y, et al. Evaluation of hepatic toxicity after repeated stereotactic body radiation therapy for recurrent hepatocellular carcinoma using deformable image registration. Sci Rep. 2018;8(1):16224. doi: 10.1038/s41598-018-34676-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13. Lee DS, Woo JY, Kim JW, Seong J. Re‐Irradiation of hepatocellular carcinoma: clinical applicability of deformable image registration. Yonsei Med J. 2016;57(1):41‐49. doi: 10.3349/ymj.2016.57.1.41 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14. Sempokuya T, Wong LL. Ten‐year survival and recurrence of hepatocellular cancer. Hepatoma Res. 2019;5. doi: 10.20517/2394-5079.2019.013 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15. McDuff SGR, Remillard KA, Zheng H, et al. Liver reirradiation for patients with hepatocellular carcinoma and liver metastasis. Pract Radiat Oncol. 2018;8(6):414‐421. doi: 10.1016/j.prro.2018.04.012 [DOI] [PubMed] [Google Scholar]
  • 16. Hall J, Moon AM, Young M, et al. Liver toxicity following reirradiation and multiple‐target SBRT for primary hepatocellular carcinoma. J Clin Oncol. 2024;42(Suppl 3):514‐514. doi: 10.1200/JCO.2024.42.3_suppl.514 [DOI] [Google Scholar]
  • 17. Huang P, Yu G, Chen J, et al. Investigation of dosimetric variations of liver radiotherapy using deformable registration of planning CT and cone‐beam CT. J Appl Clin Med Phys. 2017;18(1):66‐75. doi: 10.1002/acm2.12008 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18. Swaminath A, Massey C, Brierley JD, et al. Accumulated delivered dose response of stereotactic body radiation therapy for liver metastases. Int J Radiat Oncol. 2015;93(3):639‐648. doi: 10.1016/j.ijrobp.2015.07.2273 [DOI] [PubMed] [Google Scholar]
  • 19. Velec M, Moseley JL, Eccles CL, et al. Effect of breathing motion on radiotherapy dose accumulation in the abdomen using deformable registration. Int J Radiat Oncol. 2011;80(1):265‐272. doi: 10.1016/j.ijrobp.2010.05.023 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20. Yeo UA, Taylor ML, Supple JR, et al. Evaluation of dosimetric misrepresentations from 3D conventional planning of liver SBRT using 4D deformable dose integration. J Appl Clin Med Phys. 2014;15(6):188‐203. doi: 10.1120/jacmp.v15i6.4978 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21. Cazoulat G, Gupta AC, Al Taie MM, Koay EJ, Brock KK. Analysis and prediction of liver volume change maps derived from computational tomography scans acquired pre‐ and post‐radiation therapy. Phys Med Biol. 2023;68(20):205009. doi: 10.1088/1361-6560/acfa5f [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22. Navin PJ, Olson MC, Mendiratta‐Lala M, Hallemeier CL, Torbenson MS, Venkatesh SK. Imaging features in the liver after stereotactic body radiation therapy. RadioGraphics. 2022;42(7):2131‐2148. doi: 10.1148/rg.220084 [DOI] [PubMed] [Google Scholar]
  • 23. Murr M, Bernchou U, Bubula‐Rehm E, et al. A multi‐institutional comparison of retrospective deformable dose accumulation for online adaptive magnetic resonance‐guided radiotherapy. Phys Imaging Radiat Oncol. 2024;30. doi: 10.1016/j.phro.2024.100588 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24. Baron RL. Understanding and optimizing use of contrast material for CT of the liver. Am J Roentgenol. 1994;163(2):323‐331. doi: 10.2214/ajr.163.2.8037023 [DOI] [PubMed] [Google Scholar]
  • 25. Brock KK. Imaging and Image‐Guided Radiation Therapy in liver cancer. Semin Radiat Oncol. 2011;21(4):247‐255. doi: 10.1016/j.semradonc.2011.05.001 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26. Ichikawa T, Erturk SM, Araki T. Multiphasic contrast‐enhanced multidetector‐row CT of liver: contrast‐enhancement theory and practical scan protocol with a combination of fixed injection duration and patients’ body‐weight‐tailored dose of contrast material. Eur J Radiol. 2006;58(2):165‐176. doi: 10.1016/j.ejrad.2005.11.037 [DOI] [PubMed] [Google Scholar]
  • 27. Varian Medical Systems Inc . VELOCITY Instructions for Use. Varian Medical Systems Inc.; 2018.
  • 28. Taha AA, Hanbury A. Metrics for evaluating 3D medical image segmentation: analysis, selection, and tool. BMC Med Imaging. 2015;15:29. doi: 10.1186/s12880-015-0068-x [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29. Vaassen F, Hazelaar C, Vaniqui A, et al. Evaluation of measures for assessing time‐saving of automatic organ‐at‐risk segmentation in radiotherapy. Phys Imaging Radiat Oncol. 2020;13:1‐6. doi: 10.1016/j.phro.2019.12.001 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30. Bosma LS, Hussein M, Jameson M, et al. Tools and recommendations for commissioning and quality assurance of deformable image registration in radiotherapy. Phys Imaging Radiat Oncol. 2024;32. doi: 10.1016/j.phro.2024.100647 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31. Brock KK, Mutic S, McNutt TR, Li H, Kessler ML. Use of image registration and fusion algorithms and techniques in radiotherapy: report of the AAPM Radiation Therapy Committee Task Group No. 132. Med Phys. 2017;44(7):e43‐e76. doi: 10.1002/mp.12256 [DOI] [PubMed] [Google Scholar]
  • 32. Murr M, Brock KK, Fusella M, et al. Applicability and usage of dose mapping/accumulation in radiotherapy. Radiother Oncol J Eur Soc Ther Radiol Oncol. 2023;182. doi: 10.1016/j.radonc.2023.109527 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33. Ho J, Tumkaya T, Aryal S, Choi H, Claridge‐Chang A. Moving beyond P values: data analysis with estimation graphics. Nat Methods. 2019;16(7):565‐566. doi: 10.1038/s41592-019-0470-3 [DOI] [PubMed] [Google Scholar]
  • 34. Scheff SW. Chapter 8 – nonparametric statistics. In: Scheff SW, ed. Fundamental Statistical Principles for the Neurobiologist. Academic Press; 2016:157‐182. doi: 10.1016/B978-0-12-804753-8.00008-7 [DOI] [Google Scholar]
  • 35. Freund RJ, Wilson WJ, Mohr DL. Chapter 14 – nonparametric methods. In: Freund RJ, Wilson WJ, Mohr DL, eds. Statistical Methods. Academic Press; 2010:689‐719. doi: 10.1016/B978-0-12-374970-3.00014-7 [DOI] [Google Scholar]
  • 36. Fukumitsu N, Terunuma T, Okumura T, et al. Comparison of rigid and deformable image registration accuracy of the liver during long‐term transition after proton beam therapy. Imaging Med. 2017;9(6):149‐154. doi: 10.14303/Imaging-Medicine.1000078 [DOI] [Google Scholar]
  • 37. Cazoulat G, Anderson BM, McCulloch MM, Rigaud B, Koay EJ, Brock KK. Detection of vessel bifurcations in CT scans for automatic objective assessment of deformable image registration accuracy. Med Phys. 2021;48(10):5935‐5946. doi: 10.1002/mp.15163 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 38. Fukumitsu N, Nitta K, Terunuma T, et al. Registration error of the liver CT using deformable image registration of MIM Maestro and Velocity AI. BMC Med Imaging. 2017;17(1):30. doi: 10.1186/s12880-017-0202-z [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39. Sen A, Anderson BM, Cazoulat G, et al. Accuracy of deformable image registration techniques for alignment of longitudinal cholangiocarcinoma CT images. Med Phys. 2020;47(4):1670‐1679. doi: 10.1002/mp.14029 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40. Polan DF, Feng M, Lawrence TS, Ten Haken RK, Brock KK. Implementing radiation dose‐volume liver response in biomechanical deformable image registration. Int J Radiat Oncol. 2017;99(4):1004‐1012. doi: 10.1016/j.ijrobp.2017.06.2455 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 41. Cazoulat G, Elganainy D, Anderson BM, et al. Vasculature‐driven biomechanical deformable image registration of longitudinal liver cholangiocarcinoma computed tomographic scans. Adv Radiat Oncol. 2020;5(2):269‐278. doi: 10.1016/j.adro.2019.10.002 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 42. Yeo UJ, Taylor ML, Supple JR, Smith RL, Kron T, Franich RD. Deformable gel dosimetry II: experimental validation of DIR‐based dose‐warping. J Phys Conf Ser. 2013;444:012107. doi: 10.1088/1742-6596/444/1/012107 [DOI] [Google Scholar]
  • 43. Cheung W, Hamarneh G. n‐SIFT: n‐dimensional scale invariant feature transform. Image Process IEEE Trans On Published Online. 2009:2012‐2021. doi: 10.1109/TIP.2009.2024578 [DOI] [PubMed] [Google Scholar]
  • 44. Koay EJ, Owen D, Das P. Radiation‐induced liver disease and modern radiotherapy. Semin Radiat Oncol. 2018;28(4):321‐331. doi: 10.1016/j.semradonc.2018.06.007 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 45. Olsen CC, Welsh J, Kavanagh BD, et al. Microscopic and macroscopic tumor and parenchymal effects of liver stereotactic body radiotherapy. Int J Radiat Oncol. 2009;73(5):1414‐1424. doi: 10.1016/j.ijrobp.2008.07.032 [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplementary Materials

MP-52-0-s001.pdf (776.7KB, pdf)

Articles from Medical Physics are provided here courtesy of Wiley

RESOURCES