An International Study of Factors Affecting Variability of Dosimetry Calculations, Part 3: Contribution from Calculating Absorbed Dose from Time-Integrated Activity

DISCUSSION

This analysis focused on task 5 of the SNMMI ¹⁷⁷Lu dosimetry challenge based on 2 ¹⁷⁷Lu-DOTATATE patient datasets with the aim of acquiring a better understanding of the variability and the sources of this variability in the last step of the dosimetry workflow, conversion of TIAs to ADs. We evaluated the use of OSVs, VSVs, local deposition method, and MC simulations.

We assessed the baseline variability in AD when performed by 2 independent MPs using the same set of dosimetry methods and software. When the same corrections were applied such as mass scaling and density weighting, the differences in ADs between methods and software were low, ranging from −1% to +2% for healthy organs and from −2% to +1% for the lesions (Fig. 1). This demonstrates that consistent ADs between different experts can be obtained independently of software and method when mass scaling and similar VOIs (6,13) are used. Note that this analysis was limited to soft-tissue structures, for which the impact of tissue heterogeneity is likely to be small. We did not evaluate tissue such as lungs, bone, and (particularly) spongiosa regions relevant for bone marrow dosimetry.

A separate analysis investigated the effect on AD variability of different formats for provision of VOIs—that is, RTstruct and binary masks, selection of VOI time point, and use or nonuse of mass scaling on OSVs. The analysis was performed only for healthy organs of patient A. The lesions were not assessed, since the VOIs were defined by a radiologist on the CT of a single time point and then copied to subsequent time points, whereas for healthy organs, CT-based segmentation was performed at each scan time point.

Overall, the up to 7% variability as assessed by QCD for kidneys and spleen was due to the 2 VOI formats, the different VOIs per scan time point, and the fact that S value mass scaling was not used in all cases (Fig. 2). The largest dispersion in ADs was found for the liver, where different scenarios with and without inclusion of the liver lesions in the liver TIA were assessed (Fig. 2). For the liver, a QCD of 29% was found when all AD data were combined—that is, AD data derived from including and excluding lesions in the overall liver VOI. The findings for the liver from the baseline variability analysis were also observed in data from participants, where QCDs of up to 36% were found for most dosimetry methods and software, which can likely be attributed to averaging of lesion dose with healthy liver dose (Figs. 6 and 7). The provided VOIs in tasks 4 and 5 of the challenge did not automatically remove the lesion uptake from the liver. A region was given for the whole liver and for individual lesions, and the participants needed to decide whether and how to remove the lesions in the liver. These findings indicate a strong need for a standardization on treatment of lesion uptake in normal tissues, especially for small organs, where the effect is assumed to be larger.

There was some variability resulting from the fact that some participants used RTstruct VOIs whereas others used binary masks. The binary masks were generated from the RTstruct VOIs and had no ambiguity about the SPECT voxels to be included. However, the application of RTstructs, which are based on polygonal contours, to SPECT images to sum voxel values has some ambiguity in terms of which fraction of voxels at the edges of the VOI should be included. How this is done is an area that would seem to be a good target for standardization to remove this source of variability.

Furthermore, VOIs for the same organ can be different for the different imaging time points depending on how they are defined. It would be beneficial for the field if recommendations could be provided as to whether individual VOIs should be obtained for each separate time point or whether only one segmentation should be performed and transferred to other time points (assuming good registration).

Finally, we observed that some of the variability in ADs is caused by use or nonuse of mass scaling for OSVs. Mass scaling is essential in obtaining accurate dose values for OSVs by adapting standard-phantom S values to patient-specific organ masses and should always be performed. If VOIs are generated for each time point, we recommend using a mass of the organ that is the average of the different VOIs.

For the analysis of the submissions from participants, the local deposition method showed the largest dispersion in ADs (QCD ≤ 36%) (Fig. 6). It is likely that this finding is due to different values of the mean energies of the ¹⁷⁷Lu β-emissions used by participants (e.g., 134 vs. 147 keV (6) or not specified). This clearly requires standardization and should include the energy of the Auger electrons that are also emitted by ¹⁷⁷Lu. Smaller variability in ADs was found for the VSV method (Fig. 4). The remaining dispersion in values could potentially be attributed to the use of different sources of VSVs, the application of density weighting, or—to a lesser extent given that here we are dealing with soft tissue—whether different VSVs were used for different tissue types. Two submissions used MC simulations with different MC codes, and the results were comparable.

The common approach of OSVs showed comparable results to the VSV approach (Fig. 4). It is not surprising that the OSV approach yielded the second smallest variability for the lesions, since this approach treats lesions as isolated spheres (10–12) and spheric S values seem to be well standardized. However, differences between ADs of lesions obtained with OSVs and voxel-level approaches exist because the former accounts for cross dose as well as tumor heterogeneity.

The use of commercial dosimetry software yielded comparable results (Fig. 5), with QCDs smaller than 3% except for the liver (Fig. 7). However, QCDs of CS 2 were larger, up to 25%, suggesting that the end user choose between organ- and voxel-level dosimetry in this software package, limiting the interpretation of the results for this CS. Because of the lack of further details provided by the participants, we were unable to distinguish between organ- and voxel-level dosimetry using this software. The use of non-CS (in-house solutions or freely available) showed a larger range in ADs (Fig. 5). However, this group was quite diverse in terms of the codes, methods, and level of validation. Despite this approach, these methods still provided QCDs smaller than 7% (Fig. 6, except for liver), indicating that the choice of dosimetry method or software might not be the largest source of variability in the dosimetry workflow. Many of the freely available software are extensively validated against the literature (11,12).

Comparing voxel-level versus organ-level dosimetry approaches, we found statistically significant differences among the ADs for healthy organs, whereas the differences among lesion ADs were not statistically significant (Table 2). These findings are surprising, since one would expect differences in AD for lesions when 3D methods are chosen that are capable of accounting for cross dose, whereas organ-level approaches rely on isolated spheres as representation for lesions. In contrast, there were statistically significant differences for the spleen and kidneys, even though cross dose is included in the organ-level calculation. However, these results must be treated with caution because of the small number of submissions (n = 25) per patient and because not all submissions included results for left kidney, right kidney, and total kidney. For the comparison of CS and non-CS, statistically significant differences in ADs were found for most lesions of both patients and the normal organs of patient B except for the total kidney. No statistically significant differences were found for normal organs of patient A or for total kidney and lesions 1 and 3 of patient B. This indicates that the dosimetry software itself may not introduce statistically significant differences in AD for all patients and organs and lesions.

Generally, we can say that the overall variability as assessed by the QCD was below 7% for all dosimetry methods, software, and regions except for the liver in patient A (Figs. 6 and 7), CS 2, and the local deposition method. These findings are in concordance with previous work that addressed the differences in ADs using different methods and software (16–19).

The small number of submissions and the fact that only 2 patient cases were provided during the challenge are limitations of our study. Furthermore, the organs and lesions included in the challenge were limited to soft-tissue structures, where differences in dosimetry methods can have a smaller impact than in heterogeneous tissue. We are aware that not all possible relevant patient characteristics were covered by these 2 cases. The issue of providing VOIs in RTstruct and mask format across the multiple imaging time points is another source of variability that is specific to this challenge. This resulted in differences in ADs of up to ±20% for the different organs. We believe that this underlines the need for standardization of VOI formats and segmentation in general to avoid such differences in volumes and therefore masses.

This analysis has shown that participants made several choices within the dosimetry workflow that increased the variability of the AD results. We believe that standardization can reduce the variability of dose calculations. We therefore recommend adoption of the following best practices.

(1) For MIRD-style, organ-level dosimetry, mass scaling is essential to account for differences between the patient’s organ mass and the mass of the stylized phantoms from which S values are precalculated.

(2) For voxelized approaches other than MC dosimetry simulation, voxel density based on registered and resampled CT images should be used to take into account patient-specific tissue heterogeneities.

(3) We found variabilities in ADs of up to 36% for the liver depending on whether lesions were included or excluded from the VOIs. For ¹⁷⁷Lu, the AD from the liver excluding lesions is likely to be a better predictor of organ toxicity than if the lesions are included. We thus recommend standardization that excludes lesions present in the organ from VOIs used to estimate AD to healthy tissues using the OSV approach. 3D dosimetry methods can account for cross irradiation of lesions to healthy tissue, and VOIs for healthy tissue should consequently exclude the lesion VOI to extract mean ADs from 3D dose images.

(4) Local energy deposition methods showed high variability because different mean energies for the ¹⁷⁷Lu electron emissions were used. We recommend using a mean energy of 147.1 keV Bq⁻¹ s⁻¹ as specified in several guidelines (6,20).

(5) We observed significant variability when applying VOIs given as binary masks versus polygonal contours (as RTstruct). This could be related to differences in how polygonal masks defined at a higher resolution (e.g., using CT images) are applied to SPECT images with larger voxels and on how subvoxel contours are in general compared against binary masks. Standardization of the VOI format is needed for the application in dosimetry.

(6) It is imperative not just to report ADs but to also include a detailed description of the methods used such that if discrepancies between reported values are present, those differences can be understood and harmonization of results pursued.

(7) Widely used software, both CS and freely available, resulted in ADs that were comparable to each other. One explanation is that this software is better validated. There was greater variability in results calculated using less widely used software. Use of software that is well validated is thus highly recommended.