Does PET SUV Harmonization Affect PERCIST Response Classification?

Pre- and posttreatment PET comparative scans should ideally be obtained with identical acquisition and processing, but this is often impractical. The degree to which differing protocols affect PERCIST classification is unclear. This study evaluates the consistency of PERCIST classification across different reconstruction algorithms and whether a proprietary software tool can harmonize SUV estimation sufficiently to provide consistent response classification. Methods: Eighty-six patients with non-small cell lung cancer, colorectal liver metastases, or metastatic melanoma who were scanned for therapy monitoring purposes were prospectively recruited in this multicenter trial. Pre- and posttreatment PET scans were acquired in protocols compliant with the Society of Nuclear Medicine and Molecular Imaging and the European Association of Nuclear Medicine (EANM) acquisition guidelines and were reconstructed with a point spread function (PSF) or PSF + time-of-flight (TOF) for optimal tumor detection and also with standardized ordered-subset expectation maximization (OSEM) known to fulfill EANM harmonizing standards. After reconstruction, a proprietary software solution was applied to the PSF TOF data (PSF +/- TOF.EQ) to harmonize SUVs with the OSEM values. The impact of differing reconstructions on PERCIST classification was evaluated. Results: For the OSEMPET1/OSEMPET2 (OSEM reconstruction for pre- and post-therapeutic PET, respectively) scenario, which was taken as the reference standard, the change in SUL was -41% +/- 25 and +56% +/- 62 in the groups of tumors showing a decrease and an increase in F-18-FDG uptake, respectively. The use of PSF reconstruction affected classification of tumor response. For example, taking the PSF +/- TOFPET1/OSEMPET2 scenario increased the apparent reduction in SUL in responding tumors (-48% +/- 22) but reduced the apparent increase in SUL in progressing tumors (+37% +/- 43), as compared with the OSEMPET1/OSEMPET2, scenario. As a result, variation in reconstruction methodology (PSF +/- TOFPET1/OSEMPET2 or OSEMPET1/PSF +/- TOFPET2) led to 13 of 86 (15%) and 17 of 86 (20%) PERCIST classification discordances, respectively. Agreement was better for these scenarios with application of the propriety filter, with K values of 1 and 0.95 compared with 0.79 and 0.72, respectively. Conclusion: Reconstruction algorithm-dependent variability in PERCIST classification is a significant issue but can be overcome by harmonizing SULs using a proprietary software tool.