TY - JOUR T1 - Quantification of Lymphoma using FDG PET Heterogeneity Features for Improved Prediction of Clinical Outcome JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 594 LP - 594 VL - 59 IS - supplement 1 AU - Arman Rahmim AU - Saeed Ashrafinia AU - Steven Rowe AU - Charles Schmidtlein AU - Mikkel Vendelbo AU - Tarec El-Galaly AU - Lars Gormsen AU - Ole Munk Y1 - 2018/05/01 UR - http://jnm.snmjournals.org/content/59/supplement_1/594.abstract N2 - 594Objectives: Although current visual grading systems for response assessment on follow-up FDG PET scans in lymphomas work well, prediction of patient outcomes and information on selecting treatment derived from baseline scans is lacking. In this work, we aim to improve prediction of clinical outcome, by moving beyond conventional imaging measures, and to perform heterogeneity quantification of FDG uptake. Methods: We analyzed pre-treatment FDG PET images from 116 patients with lymphoma (73 (62.9%) males; mean age 64.1 years [SD 15.3; range 16-88]), and assessed prediction of overall survival (OS) and progression-free survival (PFS). The number of events for OS (death) and PFS (progression) were 28 and 36, respectively. First, volumetric tumor segmentations were performed: thresholding (1) 41% SUVmax, (2) SUV>2.5, and (3) SUV>3.5 (Hermes Hybrid Viewer PDR). A total of 45 features were extracted from each patient. These included (1) sex, (2) age, (3) clinical staging, (4) PET-only staging, and (5) extra-nodal involvement. We also extracted 40 quantitative imaging measures, including SUVmax, SUVmean, metabolic tumor volume (MTV) and total lesion glycolysis (TLG) (n=4), the recently introduced class of generalized effective total uptake (gETU) measures [1] which place varying degrees of emphasis on volumetric vs. uptake information (n=10), intensity histogram (n=19) and intensity-volume histogram (n=7) measures [2, 3]. All metrics were standardized according to the framework of the image biomarker standardization initiative (IBSI) [4]. Feature selection was then performed. Metrics with Pearson correlations (r) > 0.95 were considered relatively redundant, and a total of 26 features were retained. Finally, univariate and multivariate analyses were performed, which included statistical considerations (to discourage false discovery and overfitting), assessing prediction of OS and PFS. Specifically, Kaplan-Meier survival analyses were carried out, where the subjects were divided into high-risk and low-risk groups of nearly equal patients, from which the hazard ratios (HR) were computed via Cox proportional hazards regression. Results: The third segmentation method (SUV>3.5) outperformed other methods and is reported for all subsequent analysis. Table 1 depicts univariate Cox regression analysis for OS and PFS (we report on the performance of SUV, MTV, TLG, and any metric found to be significant). It is seen that even though volumetric measures (MTV and TLG) outperform SUV metrics, all these measures still perform relatively poorly for prediction of OS and PFS. By contrast, age, clinical stage, and extra-nodal involvement are stronger predictors (with respective HR values of 2.89, 2.15 and 2.62 for OS, and 2.05, 2.49 and 2.75 for PFS). By comparison, we found that two PET-based heterogeneity features, namely histogram-mode and V10, resulted in stronger predictions of outcome (HR values of 3.02 and 4.34 for OS, and 2.31 and 2.86 for PFS) [histogram-mode denotes the most common discretized gray level in histogram binning of the segmentation, while V10 quantifies the tumor volume fraction containing at least 10% of the maximum gray level (intensity)]. After correcting for multiple testing (to control for false discovery), V10 was seen to be a significant predictor of outcome for both OS and PFS, unlike all other features. Finally, multivariate analysis (Table 2) revealed that improved HR=3.64 could be obtained for PFS, when combining heterogeneity feature V10 with extra-nodal involvement in the final model. Conclusions: It was demonstrated that FDG PET heterogeneity features, in contrast to commonly invoked imaging features, hold significant potential for improved prediction of clinical outcome in patients with lymphoma. View this table:Table 1: Univariate Cox regression ([asterisk] indicates significance after correction for multiple testing) View this table:Table 2: Final multivariate Cox regression model for OS and PFS ER -