Noise-Induced Variability of Immuno-PET with Zirconium-89-Labeled Antibodies: an Analysis Based on Count-Reduced Clinical Images

Antibody imaging is of interest to improve efficacy and limit toxicity of antibody treatment by providing a predictive imaging biomarker for antibody uptake. Zr-89-PET can be used for in vivo quantification of antibody biodistribution and tumor uptake [1, 2]. Knowledge about measurement variability is required for clinical application. Usually, a test-retest study is performed for novel tracers to assess repeatability. However, for Zr-89 -immuno-PET, repeatability is unknown. A classical test-retest study design with two tracer injections is challenging in case of Zr-89 -immuno-PET because of the long half-life of Zr-89 (78.4 h). This requires more than 10 days between two injections to have less than 10 % of the radioactivity due to the first injection remaining in the body. In addition, radiation exposure is significant (0.5 mSv/MBq) [3]. To date, most clinical PET studies using Zr-89-labeled monoclonal antibodies (mAbs) are performed with an injected dose of 37 MBq, resulting in an effective dose of 18.5 mSv. We hypothesize that the relatively low signal -to- noise ratio for Zr-89 -immuno-PET acquisition (due to the low injected dose and low positron abundance of Zr-89) results in a considerable source of measurement variability. The primary objective of this study was therefore to assess noise-induced variability of quantitative uptake measures derived from Zr-89 -immuno-PET for an injected dose of 37 MBq using count-reduced images.

For this purpose, previously acquired clinical datasets can be used to assess noise-induced variability at 50 % of the original injected dose. Raw PET data (list mode data) can be split in two equal parts (Fig. 1). The split list mode data can be reconstructed into two count-reduced images [4]. Each of the two count-reduced images is considered to be a count statistically independent estimate of an image that would have been obtained with 50 % of the original injected dose for the same scan time. For example, an original image acquired with an injected dose of 74 MBq results in two count-reduced images representing an injected dose of 37 MBq (denoted as 37MBq_74inj). The count-reduced images are not independent with respect to other factors such as procedural variations or scanner drift.

In addition, we hypothesize that noise-induced variability is independent of the investigated mAb; therefore, the secondary objective was to investigate clinical datasets with three different Zr-89-labeled mAbs. Finally, the tertiary objective was to assess noise-induced variability of Zr-89 -immuno-PET for an injected dose of 18 MBq. This is the lowest injected dose used in clinical Zr-89 -immuno-PET studies for non-oncological indications, e.g., rheumatoid arthritis and multiple sclerosis [5, 6], as further limiting radiation exposure (to < 10 mSv) is necessary for these patient categories.

Knowledge of measurement variability is of interest as a small measurement error is required for the detection of even small changes over time (e.g., response evaluation, within patients). In the current study, we assessed noise-induced variability as source of measurement error (expressed as repeatability coefficient (RC) in %). For clinical relevance, assessment of reliability is important as this indicates the ability to divide patients in groups of interest despite measurement errors (e.g., response prediction, between patients). Therefore, an explorative reliability analysis was performed (expressed as intraclass correlation coefficients (ICC)) reflecting the contribution of this source of measurement variability to the observed differences (total variance) in biodistribution and tumor uptake in these datasets.

Potential clinical applications of Zr-89 -immuno-PET include use as a quantitative imaging biomarker to assess antibody uptake in normal tissue and tumor to guide individualized treatment and/or drug development [3].

Raw list mode data was retrieved for [⁸⁹Zr]antiCD20 mAb (7 patients, 20 scans), [⁸⁹Zr]antiEGFR mAb (7 patients, 21 scans), and [⁸⁹Zr]antiCD44 mAb (13 patients, 39 scans). List mode data was available for all scans at all time points, except for 1 patient in the [⁸⁹Zr]antiCD20 mAb cohort who was originally only scanned at D0 and D6, and for which no D3 scan was available.

All original datasets were split and reconstructed (Fig. 1). However, not all splits could be reconstructed due to technical limitations, leading to the exclusion of 7 out of 80 scans. Technical limitations consisted of reconstruction failure (2 scans) and missing dicom information in the original data (5 scans). Acquisition time in h p.i. (average ± SD (range min-max)) for the included data was 1.3 ± 0.5 (0.9–2.4), 72.5 ± 3.6 (65.7–77.5), and 147.6 ± 8.0 (138.9–165.7) for [⁸⁹Zr]antiCD20 mAb; 1.3 ± 0.5 (1.0–2.4), 68.4 ± 3.0 (64.6–72.8), and 143.2 ± 2.9 (138.1–145.9) for [⁸⁹Zr]antiEGFR; and 1.5 ± 0.3 (1.1–2.1), 22.1 ± 1.5 (19.9–25.0), and 97.7 ± 5.3 (91.2–114.3) for [⁸⁹Zr]antiCD44. The number of VOIs used for each analysis is denoted in Table 1, and delineated organ and tumor volumes are presented in Suppl. Table 1 (see Electronic Supplementary Material, ESM).

In this study, noise-induced variability of quantitative uptake measures was assessed for Zr-89 -immuno-PET for an injected activity of 37 MBq based on count-reduced images. As expected, a variable increase in RC was observed from D0 to D6 for the manually delineated organs (Table 1). In general, the RC for the manually delineated large organs combined (liver, spleen, kidney, lung, and brain) was within 6 % at all time points (D0, D3, D6) for [⁸⁹Zr]anti-CD20 mAb (37MBq_74inj). Larger measurement variability was observed for the bone marrow and blood pool VOIs with RC up to 40 % at D6 for [⁸⁹Zr]anti-CD20 (37 MBq_74inj). These results are in line with an increase in RC for a lower total activity in the VOI.

For tumor uptake of [⁸⁹Zr]antiCD20 mAb (37 MBq_74inj), the lowest variability was obtained for SUV_mean (27 %), followed by SUV_peak (35 %), while SUV_max (42 %) resulted in the highest measurement variability (Table 2). This is as expected, since for a given VOI the mean value takes all voxels into account, in contrast to the peak value (based on a limited number of voxels) and max value (based on a single voxel).

RC for the three Zr-89 -mAbs were similar. Therefore, the observed measurement variability is independent of the differences in biodistribution between these three Zr-89 -mAbs. These results suggest that similar noise-induced variability can be expected for other Zr-89 -mAbs, assuming harmonized image quality and a biodistribution within the same range as the Zr-89 -mAbs investigated in this study.

With the count-reduced images, noise-induced variability for an injected dose of 18 MBq was assessed for all three mAbs. RC for combined organs were up to 12 % at all time points (range D0–D6). Tumor RC varied between 20 and 54 %, depending on time point and VOI delineation method.

Despite relatively large RC for tumor, blood pool, and bone marrow, overall reliability for the three clinical Zr-89 -immuno-PET studies previously reported [7‐9] was excellent (ICCs approximately 0.9). ICC values are given to show, for these Zr-89 -mAbs in their respective patient cohorts, that the measured differences in tumor uptake do exceed the variability induced by noise. ICC values, however, cannot be extrapolated to other Zr-89 -mAbs or even different patient cohorts imaged with the same Zr-89 -mAb.

As our study provides an assessment of measurement variability due to the signal-to-noise ratio, other sources of measurement error are not represented in this reliability assessment (Eq. 2). Factors affecting 2-deoxy-2-[¹⁸F]fluoro-d-glucose ([¹⁸F]FDG)-PET quantification have been described previously [15]. Repeatability assessed by a classical test-retest study contains intra-patient variability between the test and the retest scan as well. For a test-retest study with Zr-89 -immuno-PET, the following factors are expected to play a role (1) biological factors: e.g., uptake period (higher uptake levels at increasing time interval between injection and start of PET study), estimated at 1 % for ± 1 h [16, 17] and (2) physical factors: VOI definition, estimated at < 1 % for max and peak AC to 8 % for mean [18]. Based on the results of our study, we expect that noise-induced variability will be a major contribution to measurement variability in a Zr-89 -immuno-PET test-retest study (3 % for large organs, 20 % for tumor, values presented as SD instead of RC for comparison).

In previously reported test-retest studies of repeatability of [¹⁸F]FDG-PET, RC of < 10 % have been reported for tumor uptake on [¹⁸F]FDG-PET, resulting in thresholds of 10–15 % needed to detect therapy-induced changes in patients with non-small cell lung cancer [19‐21]. For Zr-89 -immuno-PET (37 MBq_74inj), we obtained similar RC of less than 10 % for mean activity concentrations of manually delineated organs. However, for tumor uptake on Zr-89 -immuno-PET, RC up to 42 % for SUV_max reflect the much lower signal -to- noise ratio in Zr-89 -immuno-PET scans in comparison to [¹⁸F]FDG-PET scans. When immuno-PET with Zr-89-labeled mAbs is used to assess response to treatment (e.g., by alteration of antigen expression [22]), knowledge on measurement variability should be applied to set corresponding thresholds, following the example of the use of thresholds for response assessment with [¹⁸F]FDG-PET. For data already obtained, RC from our study are relevant to allow correct interpretation, as differences smaller than the measurement variability cannot be attributed to biological effects. Measurement variability (given as RC) is independent of the study population. In addition, RC of tumor and blood pool are of interest when selecting an appropriate uptake measure and/or VOI delineation method.

Tumor-to-blood ratios, which are commonly used, will result in even worse RC due to propagation of the individual RC as both numerator and denominator contain measurement variability. Future work may include an investigation to optimize delineation methods for image-derived blood pool (e.g., delineation of a larger region in the left ventricular cavity compared to a smaller fixed-size VOI in the aortic arch) to reduce noise-induced variability. We designed this study based on patient data to provide a clinically relevant assessment of measurement error. This study provides the first description of noise-induced variability for Zr-89 -immuno-PET. In literature, more sophisticated applications of the technique of count-reduced images have been described [23, 24]. By using bootstrapping, the difference between the two image estimates can be obtained more precisely, which may allow for an examination of the change of noise-induced variability with patient-specific factors such as tumor size, uptake, patient size, and location of the tumor in the patient. In addition, through application of this technique, the effect of lowering the injected dose on quantitative accuracy can be assessed.

Our results suggest that, as expected, measurement error may be dependent on volume and uptake (Suppl. Fig. 1). In future work, phantoms could be used to study dependencies of noise-induced variability on, e.g., VOI volume and activity concentration. Such phantom experiments could be combined with tumor characteristics as derived from clinical Zr-89 -immuno-PET studies (tumor volume, SUV, tumor -to- background ratio and localization) to obtain a detection limit for tumor identification. The trade-off between noise and contrast will determine the optimal time point for tumor imaging. In addition, tumor contrast on clinical images should be defined and characterized, allowing for a recommendation on the optimal time points for future clinical studies.

For potential application of immuno-PET with Zr-89-labeled antibodies as a quantitative imaging biomarker to predict which patients are likely to respond to antibody treatment, the ability to distinguish biological differences in antibody uptake between patients is required. We observed excellent ICCs (lesion-based analysis), suggesting that an injected dose of 18-37 MBq was sufficient for the datasets included. In general, justification of the injected dose should be based on the expected effect size in the population of interest.

Another clinical application of Zr-89 -immuno-PET is in vivo quantification of antibody biodistribution and tumor uptake to assess novel antibodies during early phases of drug development. For this purpose, noise-induced variability as source of measurement error as reported in our study (expressed as RC) can be applied to estimate the minimal measurement variability (e.g., for sample size calculations for novel studies).

Based on this study, noise-induced variability results in a RC for Zr-89 -immuno-PET (37 MBq) of less than 6 % for manually delineated organs combined, increasing up to 43 % at D6 for blood pool and bone marrow, assuming similar biodistribution of mAbs. The signal -to- noise ratio in Zr-89 -immuno-PET scans leads to tumor RC up to 42 %.