Why harmonization is needed when using FDG PET/CT as a prognosticator: demonstration with EARL-compliant SUV as an independent prognostic factor in lung cancer.

Abstract

To determine EARL-compliant prognostic SUV thresholds in a mature cohort of patients with locally advanced NSCLC, and to demonstrate how detrimental it is to use a threshold determined on an older-generation PET system with a newer PET/CT machine, and vice versa, or to use such a threshold with non-harmonized multicentre pooled data. This was a single-centre retrospective study including 139 consecutive stage IIIA-IIIB patients. PET data were acquired as per the EANM guidelines and reconstructed with unfiltered point spread function (PSF) reconstruction. Subsequently, a 6.3 mm Gaussian filter was applied using the EQ.PET (Siemens Healthineers) methodology to meet the EANM/EARL harmonizing standards (PSF_EARL). A multicentre study including non-EARL-compliant systems was simulated by randomly creating four groups of patients whose images were reconstructed with unfiltered PSF and PSF with Gaussian post-filtering of 3, 5, and 10 mm. Identification of optimal SUV thresholds was based on a two-fold cross-validation process that partitioned the overall sample into learning and validation subsamples. Proportional Cox hazards models were used to estimate age-adjusted and multivariable-adjusted hazard ratios (HRs) and their 95% confidence intervals. Kaplan-Meier curves were compared using the log rank test. Median follow-up was 28 months (1-104 months). For the whole population, the estimated overall survival rate at 36 months was 0.39 [0.31-0.47]. The optimal SUV_max cutoff value was 25.43 (95% CI: 23.41-26.31) and 8.47 (95% CI: 7.23-9.31) for the PSF and for the EARL-compliant dataset respectively. These SUV_max cutoff values were both significantly and independently associated with lung cancer mortality; HRs were 1.73 (1.05-2.84) and 1.92 (1.16-3.19) for the PSF and the EARL-compliant dataset respectively. When (i) applying the optimal PSF SUV_max cutoff on an EARL-compliant dataset and the optimal EARL SUV_max cutoff on a PSF dataset or (ii) applying the optimal EARL compliant SUV_max cutoff to a simulated multicentre dataset, the tumour SUV_max was no longer significantly associated with lung cancer mortality. The present study provides the PET community with an EARL-compliant SUV_max as an independent prognosticator for advanced NSCLC that should be confirmed in a larger cohort, ideally at other EARL accredited centres, and highlights the need to harmonize PET quantitative metrics when using them for risk stratification of patients.