Reply: Automated Segmentation of TMTV in DLBCL Patients: What About Method Measurement Uncertainty?

REPLY: We thank Laffon and Marthan for their interest in our work (1) and for acknowledging that bias in metabolic tumor volume (MTV) outcome is less clinically relevant than good reproducibility. We agree that estimation of the reproducibility of MTV measurement methods is important to determine measurement uncertainty. We reported that agreement between observers for assessment of MTV measurements using the same software was 91% for the method that uses 41% of maximum SUV and more than 95% for all other methods, and we considered this to be good agreement (1). The success rate of MTV measurement was unaffected by scanning conditions (whether compliant or not with the EANM Research Ltd. harmonization program) and the presence or absence of subsequent disease progression. The uptake time influenced the success rate of measurements for the method that uses 41% of maximum SUV and the method that uses majority vote 3, which were less successful with longer uptake times.

Laffon and Marthan propose that MTV cutoffs derived from PET data to guide discrimination of prognosis should be accompanied by upper and lower confidence limits based on measurement uncertainty. The main purpose of our work was not to derive cutoffs to discriminate prognosis but to take a first step to answer a methodologic question, which was to determine the optimal automatic segmentation method or methods for MTV to apply in a larger cohort. The criteria in our study focused on 2 aspects. First, did the MTV measurement methods generate plausible total tumor burden segmentations? This was prioritized over precision, as good repeatability does not necessarily provide meaningful results. Thereby, whether such (known) precision should subsequently be used to define a threshold uncertainty or gray zone is a matter of effect size in the studied population and the intended use of the biomarker. Second, to apply a method clinically or in trials, the segmentation and workflow should be fast and easy to use and have minimal observer interaction. By applying these criteria, we identified 2 candidate methods (majority vote 2 and the method based on a fixed SUV threshold of 4.0 g/mL) that can be considered for further MTV biomarker validation. For individual patient assessment to guide prognosis and when the ultimate goal is to offer personalized treatment, MTV should ideally be assessed as a continuous variable. Then, cut points and measurement errors or misclassification become less relevant.

We presented data on discriminatory power to confirm similarity for the different segmentation methods as shown previously (2) and to support the argument that choice of method can be based on ease of use and success rates in giving plausible volumes under various conditions. For the current study, we used a case-control design to test parameters that might influence the best segmentation method—meaning that the patient population and any derived cutoffs would not be representative of usual clinical practice. We are progressing with MTV measurement in a large warehouse of clinical and scan data in patients with non-Hodgkin lymphoma (https://petralymphoma.org/). Sufficient data are required to derive robust optimal MTV cutoffs for training, validation, and test datasets. In these studies, measurement error, confidence limits, and uncertainty will be considered.

Finally, MTV is a robust predictor of prognosis in diffuse large B-cell lymphoma but will likely need to be factored into an algorithm with baseline clinical factors, including the international prognostic index (3), and potentially with emerging biomarkers that reflect tumor dissemination and molecular heterogeneity (4,5) and dynamic response markers (3,4).

• © 2021 by the Society of Nuclear Medicine and Molecular Imaging.